.github/pull_request_template.md

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

CMakeLists.txt

@@ -72,31 +72,37 @@ option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF
 if(SD_CUDA)
     message("-- Use CUDA as backend stable-diffusion")
     set(GGML_CUDA ON)
+    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_METAL)
     message("-- Use Metal as backend stable-diffusion")
     set(GGML_METAL ON)
+    add_definitions(-DSD_USE_METAL)
 endif()
 
 if (SD_VULKAN)
     message("-- Use Vulkan as backend stable-diffusion")
     set(GGML_VULKAN ON)
+    add_definitions(-DSD_USE_VULKAN)
 endif ()
 
 if (SD_OPENCL)
     message("-- Use OpenCL as backend stable-diffusion")
     set(GGML_OPENCL ON)
+    add_definitions(-DSD_USE_OPENCL)
 endif ()
 
 if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
+    add_definitions(-DSD_USE_CUDA)
 endif ()
 
 if(SD_MUSA)
     message("-- Use MUSA as backend stable-diffusion")
     set(GGML_MUSA ON)
+    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_WEBP)
@@ -106,8 +112,7 @@ if(SD_WEBP)
           "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
     endif()
     if(SD_USE_SYSTEM_WEBP)
-        find_package(WebP)
+        find_package(WebP REQUIRED)
-        if(WebP_FOUND)
         add_library(webp ALIAS WebP::webp)
         # libwebp CMake target naming is not consistent across versions/distros.
         # Some export WebP::libwebpmux, others export WebP::webpmux.
@@ -121,14 +126,6 @@ if(SD_WEBP)
                 "Expected WebP::libwebpmux or WebP::webpmux."
             )
         endif()
-        else()
-            find_package(PkgConfig REQUIRED)
-            pkg_check_modules(WebP REQUIRED IMPORTED_TARGET GLOBAL libwebp)
-            pkg_check_modules(WebPMux REQUIRED IMPORTED_TARGET GLOBAL libwebpmux)
-            link_libraries(PkgConfig::WebP)
-            link_libraries(PkgConfig::WebPMux)
-            add_library(libwebpmux ALIAS PkgConfig::WebPMux)
-        endif()
     endif()
 endif()
 
@@ -142,13 +139,6 @@ if(SD_WEBM)
           "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
     endif()
     if(SD_USE_SYSTEM_WEBM)
-        find_package(PkgConfig)
-        if(PkgConfig_FOUND)
-            pkg_check_modules(WebM REQUIRED IMPORTED_TARGET GLOBAL libwebm)
-        endif()
-        if(PkgConfig_FOUND AND WebM_FOUND)
-            link_libraries(PkgConfig::WebM)
-        else()
         find_path(WEBM_INCLUDE_DIR
             NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
             PATH_SUFFIXES webm
@@ -163,7 +153,6 @@ if(SD_WEBM)
             INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
     endif()
 endif()
-endif()
 
 set(SD_LIB stable-diffusion)
 
@@ -233,6 +222,7 @@ if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+    add_definitions(-DSD_USE_SYCL)
     # disable fast-math on host, see:
     # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
     if (WIN32)

CONTRIBUTING.md

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

README.md

@@ -58,7 +58,6 @@ API and command-line option may change frequently.***
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
-    - [HiDream-O1-Image](./docs/hidream_o1_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -78,10 +77,9 @@ API and command-line option may change frequently.***
   - OpenCL
   - SYCL
 - Supported weight formats
-  - Pytorch checkpoint (`.ckpt` or `.pth` or `.pt`)
+  - Pytorch checkpoint (`.ckpt` or `.pth`)
   - Safetensors (`.safetensors`)
   - GGUF (`.gguf`)
-- Convert mode supports converting model weights to `.gguf` or `.safetensors`
 - Supported platforms
     - Linux
     - Mac OS
@@ -149,7 +147,6 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [Ovis-Image](./docs/ovis_image.md)
 - [Anima](./docs/anima.md)
 - [ERNIE-Image](./docs/ernie_image.md)
-- [HiDream-O1-Image](./docs/hidream_o1_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -165,7 +162,6 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 
 * Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
-* Golang (non-cgo): [l8bloom/gosd](https://github.com/l8bloom/gosd)
 * C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
 * Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
 * Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)

docs/caching.md

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

docs/hidream_o1_image.md

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

examples/cli/README.md

@@ -4,17 +4,14 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>         path to write result image to. you can use printf-style %d format specifiers for image
+  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
-                                sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs
+                              ./output.png) (eg. output_%03d.png). For video generation, single-file outputs support .avi, .webm, and animated .webp
-                                support .avi, .webm, and animated .webp
+  --preview-path <string>     path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp
+  --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
+                              every step)
+  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
   --image <string>            path to the image to inspect (for metadata mode)
   --metadata-format <string>  metadata output format, one of [text, json] (default: text)
-  --preview-path <string>       path to write preview image to (default: ./preview.png). Multi-frame previews support
-                                .avi, .webm, and animated .webp
-  --preview-interval <int>      interval in denoising steps between consecutive updates of the image preview file
-                                (default is 1, meaning updating at every step)
-  --output-begin-idx <int>      starting index for output image sequence, must be non-negative (default 0 if specified
-                                %d in output path, 1 otherwise)
   --canny                     apply canny preprocessor (edge detection)
   --convert-name              convert tensor name (for convert mode)
   -v, --verbose               print extra info
@@ -34,8 +31,7 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
-                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -47,18 +43,16 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
-  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           then threads will be set to the number of CPU physical cores
+                                           CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-                                           graph splitting
+  --vae-tiling                             process vae in tiles to reduce memory usage
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
-  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
-                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -73,19 +67,20 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           q4_K). If not specified, the default is the type of the weight file
+                                           type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           immediately will be used.The immediately mode may have precision and
-                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           immediately mode may have precision and compatibility issues with quantized
+                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           other hand, is exactly the opposite.
-                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-                                           opposite.
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
+                                           (overrides --vae-tile-size)
 
 Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -94,101 +89,69 @@ Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           inside should be stored as images in lexicographical (character) order. For
+                                           lexicographical (character) order. For example, if the control video path is
-                                           example, if the control video path is `frames`, the directory contain images
+                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
-                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
-  --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
-                                           (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
-                                           antialiased), or a model name under --hires-upscalers-dir (default: Latent)
-  --extra-sample-args <string>             extra sampler args, key=value list. Currently lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
+                                           will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
+                                           NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
-  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
-  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
-  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
-  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
-                                           128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
-                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           disabled, a value of 2.5 is nice for sd3.5 medium
+                                           medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
-                                           res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
-                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
-                                           (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
-                                           0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
-                                           res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
-  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
-                                           destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
-                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
-  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
-  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
-  --increase-ref-index                     automatically increase the indices of references images based on the order
-                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
-  --vae-tiling                             process vae in tiles to reduce memory usage
-  --hires                                  enable highres fix
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
-  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
+                                           tcd, res_multistep, res_2s, er_sde] (default: euler for Flux/SD3/Wan, euler_a
-                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           otherwise)
-  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
+                                           ddim_trailing, tcd, res_multistep, res_2s, er_sde] default: euler for Flux/SD3/Wan,
-                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           euler_a otherwise
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
+                                           kl_optimal, lcm, bong_tangent], default: discrete
-                                           discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
-                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level),
-                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
-                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=;
-                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples:
-                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+                                           "threshold=0.25" or "threshold=1.5,reset=0" or "w=0.4,window=2"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
-                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
-                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```
 
 Metadata mode inspects PNG/JPEG container metadata without loading any model:

examples/cli/main.cpp

@@ -278,9 +278,7 @@ void parse_args(int argc, const char** argv, SDCliParams& cli_params, SDContextP
     bool valid = cli_params.resolve_and_validate();
     if (valid && cli_params.mode != METADATA) {
         valid = ctx_params.resolve_and_validate(cli_params.mode) &&
-                gen_params.resolve_and_validate(cli_params.mode,
+                gen_params.resolve_and_validate(cli_params.mode, ctx_params.lora_model_dir);
-                                                ctx_params.lora_model_dir,
-                                                ctx_params.hires_upscalers_dir);
     }
 
     if (!valid) {
@@ -433,9 +431,8 @@ bool save_results(const SDCliParams& cli_params,
         if (!img.data)
             return false;
 
-        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx;
         std::string params = gen_params.embed_image_metadata
-                                          ? get_image_params(ctx_params, gen_params, metadata_seed, cli_params.mode)
+                                 ? get_image_params(ctx_params, gen_params, gen_params.seed + idx)
                                  : "";
         const bool ok      = write_image_to_file(path.string(), img.data, img.width, img.height, img.channel, params, 90);
         LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
@@ -691,13 +688,6 @@ int main(int argc, const char* argv[]) {
         vae_decode_only = false;
     }
 
-    if (gen_params.hires_enabled &&
-        (gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_MODEL ||
-         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_LANCZOS ||
-         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_NEAREST)) {
-        vae_decode_only = false;
-    }
-
     sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, true, cli_params.taesd_preview);
 
     SDImageVec results;

examples/common/common.cpp

@@ -107,60 +107,47 @@ static bool is_absolute_path(const std::string& p) {
 
 std::string ArgOptions::wrap_text(const std::string& text, size_t width, size_t indent) {
     std::ostringstream oss;
-    size_t pos      = 0;
     size_t line_len = 0;
+    size_t pos      = 0;
 
     while (pos < text.size()) {
+        // Preserve manual newlines
         if (text[pos] == '\n') {
             oss << '\n'
                 << std::string(indent, ' ');
-            line_len = 0;
+            line_len = indent;
             ++pos;
             continue;
         }
 
-        if (std::isspace(static_cast<unsigned char>(text[pos]))) {
+        // Add the character
-            ++pos;
+        oss << text[pos];
-            continue;
-        }
-
-        size_t word_start = pos;
-        while (pos < text.size() &&
-               text[pos] != '\n' &&
-               !std::isspace(static_cast<unsigned char>(text[pos]))) {
-            ++pos;
-        }
-
-        std::string word = text.substr(word_start, pos - word_start);
-        while (!word.empty()) {
-            size_t separator_len = line_len == 0 ? 0 : 1;
-            if (line_len + separator_len + word.size() <= width) {
-                if (separator_len > 0) {
-                    oss << ' ';
         ++line_len;
-                }
+        ++pos;
-                oss << word;
-                line_len += word.size();
-                word.clear();
-                continue;
-            }
 
-            if (line_len > 0) {
+        // If the current line exceeds width, try to break at the last space
-                oss << '\n'
+        if (line_len >= width) {
-                    << std::string(indent, ' ');
+            std::string current = oss.str();
-                line_len = 0;
+            size_t back         = current.size();
-                continue;
-            }
 
-            size_t chunk_len = std::min(width, word.size());
+            // Find the last space (for a clean break)
-            oss << word.substr(0, chunk_len);
+            while (back > 0 && current[back - 1] != ' ' && current[back - 1] != '\n')
-            line_len = chunk_len;
+                --back;
-            word.erase(0, chunk_len);
+
-            if (!word.empty()) {
+            // If found a space to break on
-                oss << '\n'
+            if (back > 0 && current[back - 1] != '\n') {
+                std::string before = current.substr(0, back - 1);
+                std::string after  = current.substr(back);
+                oss.str("");
+                oss.clear();
+                oss << before << "\n"
+                    << std::string(indent, ' ') << after;
+            } else {
+                // If no space found, just break at width
+                oss << "\n"
                     << std::string(indent, ' ');
-                line_len = 0;
             }
+            line_len = indent;
         }
     }
 
@@ -364,10 +351,7 @@ ArgOptions SDContextParams::get_options() {
          "--lora-model-dir",
          "lora model directory",
          &lora_model_dir},
-        {"",
+
-         "--hires-upscalers-dir",
-         "highres fix upscaler model directory",
-         &hires_upscalers_dir},
         {"",
          "--tensor-type-rules",
          "weight type per tensor pattern (example: \"^vae\\.=f16,model\\.=q8_0\")",
@@ -394,12 +378,7 @@ ArgOptions SDContextParams::get_options() {
          &chroma_t5_mask_pad},
     };
 
-    options.float_options = {
+    options.float_options = {};
-        {"",
-         "--max-vram",
-         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting",
-         &max_vram},
-    };
 
     options.bool_options = {
         {"",
@@ -670,12 +649,10 @@ std::string SDContextParams::to_string() const {
         << "  wtype: " << sd_type_name(wtype) << ",\n"
         << "  tensor_type_rules: \"" << tensor_type_rules << "\",\n"
         << "  lora_model_dir: \"" << lora_model_dir << "\",\n"
-        << "  hires_upscalers_dir: \"" << hires_upscalers_dir << "\",\n"
         << "  photo_maker_path: \"" << photo_maker_path << "\",\n"
         << "  rng_type: " << sd_rng_type_name(rng_type) << ",\n"
         << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
         << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
-        << "  max_vram: " << max_vram << ",\n"
         << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
         << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
         << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
@@ -750,7 +727,6 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
         chroma_use_t5_mask,
         chroma_t5_mask_pad,
         qwen_image_zero_cond_t,
-        max_vram,
     };
     return sd_ctx_params;
 }
@@ -801,16 +777,6 @@ ArgOptions SDGenerationParams::get_options() {
          "--pm-id-embed-path",
          "path to PHOTOMAKER v2 id embed",
          &pm_id_embed_path},
-        {"",
-         "--hires-upscaler",
-         "highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent (nearest-exact), "
-         "Latent (antialiased), Latent (bicubic), Latent (bicubic antialiased), or a model name "
-         "under --hires-upscalers-dir (default: Latent)",
-         &hires_upscaler},
-        {"",
-         "--extra-sample-args",
-         "extra sampler args, key=value list. Currently lcm supports noise_clip_std, noise_scale_start, noise_scale_end",
-         &extra_sample_args},
     };
 
     options.int_options = {
@@ -860,22 +826,6 @@ ArgOptions SDGenerationParams::get_options() {
          "--upscale-tile-size",
          "tile size for ESRGAN upscaling (default: 128)",
          &upscale_tile_size},
-        {"",
-         "--hires-width",
-         "highres fix target width, 0 to use --hires-scale (default: 0)",
-         &hires_width},
-        {"",
-         "--hires-height",
-         "highres fix target height, 0 to use --hires-scale (default: 0)",
-         &hires_height},
-        {"",
-         "--hires-steps",
-         "highres fix second pass sample steps, 0 to reuse --steps (default: 0)",
-         &hires_steps},
-        {"",
-         "--hires-upscale-tile-size",
-         "highres fix upscaler tile size, reserved for model-backed upscalers (default: 128)",
-         &hires_upscale_tile_size},
     };
 
     options.float_options = {
@@ -963,14 +913,6 @@ ArgOptions SDGenerationParams::get_options() {
          "--vae-tile-overlap",
          "tile overlap for vae tiling, in fraction of tile size (default: 0.5)",
          &vae_tiling_params.target_overlap},
-        {"",
-         "--hires-scale",
-         "highres fix scale when target size is not set (default: 2.0)",
-         &hires_scale},
-        {"",
-         "--hires-denoising-strength",
-         "highres fix second pass denoising strength (default: 0.7)",
-         &hires_denoising_strength},
     };
 
     options.bool_options = {
@@ -994,11 +936,6 @@ ArgOptions SDGenerationParams::get_options() {
          "process vae in tiles to reduce memory usage",
          true,
          &vae_tiling_params.enabled},
-        {"",
-         "--hires",
-         "enable highres fix",
-         true,
-         &hires_enabled},
     };
 
     auto on_seed_arg = [&](int argc, const char** argv, int index) {
@@ -1248,12 +1185,12 @@ ArgOptions SDGenerationParams::get_options() {
          on_seed_arg},
         {"",
          "--sampling-method",
-         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
+         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde] "
          "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
          on_sample_method_arg},
         {"",
          "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
+         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde]"
          " default: euler for Flux/SD3/Wan, euler_a otherwise",
          on_high_noise_sample_method_arg},
         {"",
@@ -1487,37 +1424,6 @@ static bool parse_lora_json_field(const json& parent,
     return true;
 }
 
-static bool resolve_model_file_from_dir(const std::string& model_name,
-                                        const std::string& model_dir,
-                                        const std::vector<std::string>& valid_ext,
-                                        const char* label,
-                                        std::string& resolved_path) {
-    if (model_dir.empty()) {
-        LOG_ERROR("%s directory is empty", label);
-        return false;
-    }
-    if (model_name.empty() ||
-        model_name.find('/') != std::string::npos ||
-        model_name.find('\\') != std::string::npos ||
-        fs::path(model_name).has_root_path() ||
-        fs::path(model_name).has_extension()) {
-        LOG_ERROR("%s must be a model name without path or extension: %s", label, model_name.c_str());
-        return false;
-    }
-
-    fs::path model_dir_path = model_dir;
-    for (const auto& ext : valid_ext) {
-        fs::path try_path = model_dir_path / (model_name + ext);
-        if (fs::exists(try_path) && fs::is_regular_file(try_path)) {
-            resolved_path = try_path.lexically_normal().string();
-            return true;
-        }
-    }
-
-    LOG_ERROR("can not find %s %s in %s", label, model_name.c_str(), model_dir_path.lexically_normal().string().c_str());
-    return false;
-}
-
 bool SDGenerationParams::from_json_str(
     const std::string& json_str,
     const std::function<std::string(const std::string&)>& lora_path_resolver) {
@@ -1581,37 +1487,8 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("increase_ref_index", increase_ref_index);
     load_if_exists("embed_image_metadata", embed_image_metadata);
 
-    if (j.contains("hires") && j["hires"].is_object()) {
-        const json& hires_json = j["hires"];
-        if (hires_json.contains("enabled") && hires_json["enabled"].is_boolean()) {
-            hires_enabled = hires_json["enabled"];
-        }
-        if (hires_json.contains("upscaler") && hires_json["upscaler"].is_string()) {
-            hires_upscaler = hires_json["upscaler"];
-        }
-        if (hires_json.contains("scale") && hires_json["scale"].is_number()) {
-            hires_scale = hires_json["scale"];
-        }
-        if (hires_json.contains("target_width") && hires_json["target_width"].is_number_integer()) {
-            hires_width = hires_json["target_width"];
-        }
-        if (hires_json.contains("target_height") && hires_json["target_height"].is_number_integer()) {
-            hires_height = hires_json["target_height"];
-        }
-        if (hires_json.contains("steps") && hires_json["steps"].is_number_integer()) {
-            hires_steps = hires_json["steps"];
-        }
-        if (hires_json.contains("denoising_strength") && hires_json["denoising_strength"].is_number()) {
-            hires_denoising_strength = hires_json["denoising_strength"];
-        }
-        if (hires_json.contains("upscale_tile_size") && hires_json["upscale_tile_size"].is_number_integer()) {
-            hires_upscale_tile_size = hires_json["upscale_tile_size"];
-        }
-    }
-
     auto parse_sample_params_json = [&](const json& sample_json,
                                         sd_sample_params_t& target_params,
-                                        std::string& target_extra_sample_args,
                                         std::vector<int>& target_skip_layers,
                                         std::vector<float>* target_custom_sigmas) {
         if (sample_json.contains("sample_steps") && sample_json["sample_steps"].is_number_integer()) {
@@ -1626,9 +1503,6 @@ bool SDGenerationParams::from_json_str(
         if (sample_json.contains("flow_shift") && sample_json["flow_shift"].is_number()) {
             target_params.flow_shift = sample_json["flow_shift"];
         }
-        if (sample_json.contains("extra_sample_args") && sample_json["extra_sample_args"].is_string()) {
-            target_extra_sample_args = sample_json["extra_sample_args"].get<std::string>();
-        }
         if (target_custom_sigmas != nullptr &&
             sample_json.contains("custom_sigmas") &&
             sample_json["custom_sigmas"].is_array()) {
@@ -1676,12 +1550,11 @@ bool SDGenerationParams::from_json_str(
     };
 
     if (j.contains("sample_params") && j["sample_params"].is_object()) {
-        parse_sample_params_json(j["sample_params"], sample_params, extra_sample_args, skip_layers, &custom_sigmas);
+        parse_sample_params_json(j["sample_params"], sample_params, skip_layers, &custom_sigmas);
     }
     if (j.contains("high_noise_sample_params") && j["high_noise_sample_params"].is_object()) {
         parse_sample_params_json(j["high_noise_sample_params"],
                                  high_noise_sample_params,
-                                 high_noise_extra_sample_args,
                                  high_noise_skip_layers,
                                  nullptr);
     }
@@ -1927,7 +1800,7 @@ bool SDGenerationParams::initialize_cache_params() {
     return true;
 }
 
-bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict) {
+bool SDGenerationParams::resolve(const std::string& lora_model_dir, bool strict) {
     if (high_noise_sample_params.sample_steps <= 0) {
         high_noise_sample_params.sample_steps = -1;
     }
@@ -1946,27 +1819,6 @@ bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::s
         sample_params.sample_steps = std::clamp(sample_params.sample_steps, 1, 100);
     }
 
-    hires_upscaler_model_path.clear();
-    if (hires_enabled) {
-        if (hires_upscaler.empty()) {
-            hires_upscaler = "Latent";
-        }
-        resolved_hires_upscaler = str_to_sd_hires_upscaler(hires_upscaler.c_str());
-        if (resolved_hires_upscaler == SD_HIRES_UPSCALER_NONE) {
-            hires_enabled = false;
-        } else if (resolved_hires_upscaler == SD_HIRES_UPSCALER_COUNT) {
-            static const std::vector<std::string> valid_ext = {".gguf", ".safetensors", ".pt", ".pth"};
-            if (!resolve_model_file_from_dir(hires_upscaler,
-                                             hires_upscalers_dir,
-                                             valid_ext,
-                                             "hires upscaler",
-                                             hires_upscaler_model_path)) {
-                return false;
-            }
-            resolved_hires_upscaler = SD_HIRES_UPSCALER_MODEL;
-        }
-    }
-
     prompt_with_lora = prompt;
     if (!lora_model_dir.empty()) {
         extract_and_remove_lora(lora_model_dir);
@@ -2031,29 +1883,6 @@ bool SDGenerationParams::validate(SDMode mode) {
         return false;
     }
 
-    if (hires_enabled) {
-        if (hires_width < 0 || hires_height < 0) {
-            LOG_ERROR("error: hires target width and height must be >= 0");
-            return false;
-        }
-        if (hires_scale <= 0.f && hires_width <= 0 && hires_height <= 0) {
-            LOG_ERROR("error: hires scale must be positive when target size is not set");
-            return false;
-        }
-        if (hires_steps < 0) {
-            LOG_ERROR("error: hires steps must be >= 0");
-            return false;
-        }
-        if (hires_denoising_strength <= 0.f || hires_denoising_strength > 1.f) {
-            LOG_ERROR("error: hires denoising strength must be in (0.0, 1.0]");
-            return false;
-        }
-        if (hires_upscale_tile_size < 1) {
-            LOG_ERROR("error: hires upscale tile size must be positive");
-            return false;
-        }
-    }
-
     if (mode == UPSCALE) {
         if (init_image_path.length() == 0) {
             LOG_ERROR("error: upscale mode needs an init image (--init-img)\n");
@@ -2064,11 +1893,8 @@ bool SDGenerationParams::validate(SDMode mode) {
     return true;
 }
 
-bool SDGenerationParams::resolve_and_validate(SDMode mode,
+bool SDGenerationParams::resolve_and_validate(SDMode mode, const std::string& lora_model_dir, bool strict) {
-                                              const std::string& lora_model_dir,
+    if (!resolve(lora_model_dir, strict)) {
-                                              const std::string& hires_upscalers_dir,
-                                              bool strict) {
-    if (!resolve(lora_model_dir, hires_upscalers_dir, strict)) {
         return false;
     }
     if (!validate(mode)) {
@@ -2108,8 +1934,6 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
     sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
-    sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
-    high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     sd_pm_params_t pm_params = {
@@ -2141,16 +1965,6 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.pm_params             = pm_params;
     params.vae_tiling_params     = vae_tiling_params;
     params.cache                 = cache_params;
-
-    params.hires.enabled            = hires_enabled;
-    params.hires.upscaler           = resolved_hires_upscaler;
-    params.hires.model_path         = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
-    params.hires.scale              = hires_scale;
-    params.hires.target_width       = hires_width;
-    params.hires.target_height      = hires_height;
-    params.hires.steps              = hires_steps;
-    params.hires.denoising_strength = hires_denoising_strength;
-    params.hires.upscale_tile_size  = hires_upscale_tile_size;
     return params;
 }
 
@@ -2179,8 +1993,6 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
     sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
-    sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
-    high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     params.loras                    = lora_vec.empty() ? nullptr : lora_vec.data();
@@ -2277,15 +2089,6 @@ std::string SDGenerationParams::to_string() const {
         << "  seed: " << seed << ",\n"
         << "  upscale_repeats: " << upscale_repeats << ",\n"
         << "  upscale_tile_size: " << upscale_tile_size << ",\n"
-        << "  hires: { enabled: " << (hires_enabled ? "true" : "false")
-        << ", upscaler: \"" << hires_upscaler << "\""
-        << ", model_path: \"" << hires_upscaler_model_path << "\""
-        << ", scale: " << hires_scale
-        << ", target_width: " << hires_width
-        << ", target_height: " << hires_height
-        << ", steps: " << hires_steps
-        << ", denoising_strength: " << hires_denoising_strength
-        << ", upscale_tile_size: " << hires_upscale_tile_size << " },\n"
         << "  vae_tiling_params: { "
         << vae_tiling_params.enabled << ", "
         << vae_tiling_params.tile_size_x << ", "
@@ -2301,193 +2104,7 @@ std::string version_string() {
     return std::string("stable-diffusion.cpp version ") + sd_version() + ", commit " + sd_commit();
 }
 
-static std::string safe_json_string(const char* value) {
+std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed) {
-    return value ? value : "";
-}
-
-static void set_json_basename_if_not_empty(json& target, const char* key, const std::string& path) {
-    if (!path.empty()) {
-        target[key] = sd_basename(path);
-    }
-}
-
-static json build_sampling_metadata_json(const sd_sample_params_t& sample_params,
-                                         const std::vector<int>& skip_layers,
-                                         const std::vector<float>* custom_sigmas = nullptr) {
-    json sampling = {
-        {"steps", sample_params.sample_steps},
-        {"eta", sample_params.eta},
-        {"shifted_timestep", sample_params.shifted_timestep},
-        {"flow_shift", sample_params.flow_shift},
-        {"extra_sample_args", safe_json_string(sample_params.extra_sample_args)},
-        {"guidance",
-         {
-             {"txt_cfg", sample_params.guidance.txt_cfg},
-             {"img_cfg", sample_params.guidance.img_cfg},
-             {"distilled_guidance", sample_params.guidance.distilled_guidance},
-             {"slg",
-              {
-                  {"scale", sample_params.guidance.slg.scale},
-                  {"layers", skip_layers},
-                  {"start", sample_params.guidance.slg.layer_start},
-                  {"end", sample_params.guidance.slg.layer_end},
-              }},
-         }},
-    };
-    if (sample_params.sample_method != SAMPLE_METHOD_COUNT) {
-        sampling["method"] = safe_json_string(sd_sample_method_name(sample_params.sample_method));
-    }
-    if (sample_params.scheduler != SCHEDULER_COUNT) {
-        sampling["scheduler"] = safe_json_string(sd_scheduler_name(sample_params.scheduler));
-    }
-    if (custom_sigmas != nullptr) {
-        sampling["custom_sigmas"] = *custom_sigmas;
-    }
-    return sampling;
-}
-
-std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
-                                            const SDGenerationParams& gen_params,
-                                            int64_t seed,
-                                            SDMode mode) {
-    json root;
-    root["schema"]    = "sdcpp.image.params/v1";
-    root["mode"]      = mode == VID_GEN ? "vid_gen" : "img_gen";
-    root["generator"] = {
-        {"name", "stable-diffusion.cpp"},
-        {"version", safe_json_string(sd_version())},
-        {"commit", safe_json_string(sd_commit())},
-    };
-    root["seed"]   = seed;
-    root["width"]  = gen_params.get_resolved_width();
-    root["height"] = gen_params.get_resolved_height();
-
-    root["prompt"] = {
-        {"positive", gen_params.prompt},
-        {"negative", gen_params.negative_prompt},
-    };
-    root["sampling"] = build_sampling_metadata_json(gen_params.sample_params,
-                                                    gen_params.skip_layers,
-                                                    &gen_params.custom_sigmas);
-
-    json models;
-    set_json_basename_if_not_empty(models, "model", ctx_params.model_path);
-    set_json_basename_if_not_empty(models, "clip_l", ctx_params.clip_l_path);
-    set_json_basename_if_not_empty(models, "clip_g", ctx_params.clip_g_path);
-    set_json_basename_if_not_empty(models, "clip_vision", ctx_params.clip_vision_path);
-    set_json_basename_if_not_empty(models, "t5xxl", ctx_params.t5xxl_path);
-    set_json_basename_if_not_empty(models, "llm", ctx_params.llm_path);
-    set_json_basename_if_not_empty(models, "llm_vision", ctx_params.llm_vision_path);
-    set_json_basename_if_not_empty(models, "diffusion_model", ctx_params.diffusion_model_path);
-    set_json_basename_if_not_empty(models, "high_noise_diffusion_model", ctx_params.high_noise_diffusion_model_path);
-    set_json_basename_if_not_empty(models, "vae", ctx_params.vae_path);
-    set_json_basename_if_not_empty(models, "taesd", ctx_params.taesd_path);
-    set_json_basename_if_not_empty(models, "control_net", ctx_params.control_net_path);
-    root["models"] = std::move(models);
-
-    root["clip_skip"]             = gen_params.clip_skip;
-    root["strength"]              = gen_params.strength;
-    root["control_strength"]      = gen_params.control_strength;
-    root["auto_resize_ref_image"] = gen_params.auto_resize_ref_image;
-    root["increase_ref_index"]    = gen_params.increase_ref_index;
-    if (mode == VID_GEN) {
-        root["video"] = {
-            {"frame_count", gen_params.video_frames},
-            {"fps", gen_params.fps},
-        };
-        root["moe_boundary"]        = gen_params.moe_boundary;
-        root["vace_strength"]       = gen_params.vace_strength;
-        root["high_noise_sampling"] = build_sampling_metadata_json(gen_params.high_noise_sample_params,
-                                                                   gen_params.high_noise_skip_layers);
-    }
-
-    root["rng"] = safe_json_string(sd_rng_type_name(ctx_params.rng_type));
-    if (ctx_params.sampler_rng_type != RNG_TYPE_COUNT) {
-        root["sampler_rng"] = safe_json_string(sd_rng_type_name(ctx_params.sampler_rng_type));
-    }
-
-    json loras = json::array();
-    for (const auto& entry : gen_params.lora_map) {
-        loras.push_back({
-            {"name", sd_basename(entry.first)},
-            {"multiplier", entry.second},
-            {"is_high_noise", false},
-        });
-    }
-    for (const auto& entry : gen_params.high_noise_lora_map) {
-        loras.push_back({
-            {"name", sd_basename(entry.first)},
-            {"multiplier", entry.second},
-            {"is_high_noise", true},
-        });
-    }
-    if (!loras.empty()) {
-        root["loras"] = std::move(loras);
-    }
-
-    if (gen_params.hires_enabled) {
-        root["hires"] = {
-            {"enabled", gen_params.hires_enabled},
-            {"upscaler", gen_params.hires_upscaler},
-            {"model", gen_params.hires_upscaler_model_path.empty() ? "" : sd_basename(gen_params.hires_upscaler_model_path)},
-            {"scale", gen_params.hires_scale},
-            {"target_width", gen_params.hires_width},
-            {"target_height", gen_params.hires_height},
-            {"steps", gen_params.hires_steps},
-            {"denoising_strength", gen_params.hires_denoising_strength},
-            {"upscale_tile_size", gen_params.hires_upscale_tile_size},
-        };
-    }
-
-    if (gen_params.cache_params.mode != SD_CACHE_DISABLED) {
-        root["cache"] = {
-            {"requested_mode", gen_params.cache_mode},
-            {"requested_option", gen_params.cache_option},
-            {"mode", gen_params.cache_params.mode},
-            {"scm_mask", gen_params.scm_mask},
-            {"scm_policy_dynamic", gen_params.scm_policy_dynamic},
-            {"reuse_threshold", gen_params.cache_params.reuse_threshold},
-            {"start_percent", gen_params.cache_params.start_percent},
-            {"end_percent", gen_params.cache_params.end_percent},
-            {"error_decay_rate", gen_params.cache_params.error_decay_rate},
-            {"use_relative_threshold", gen_params.cache_params.use_relative_threshold},
-            {"reset_error_on_compute", gen_params.cache_params.reset_error_on_compute},
-            {"Fn_compute_blocks", gen_params.cache_params.Fn_compute_blocks},
-            {"Bn_compute_blocks", gen_params.cache_params.Bn_compute_blocks},
-            {"residual_diff_threshold", gen_params.cache_params.residual_diff_threshold},
-            {"max_warmup_steps", gen_params.cache_params.max_warmup_steps},
-            {"max_cached_steps", gen_params.cache_params.max_cached_steps},
-            {"max_continuous_cached_steps", gen_params.cache_params.max_continuous_cached_steps},
-            {"taylorseer_n_derivatives", gen_params.cache_params.taylorseer_n_derivatives},
-            {"taylorseer_skip_interval", gen_params.cache_params.taylorseer_skip_interval},
-            {"spectrum_w", gen_params.cache_params.spectrum_w},
-            {"spectrum_m", gen_params.cache_params.spectrum_m},
-            {"spectrum_lam", gen_params.cache_params.spectrum_lam},
-            {"spectrum_window_size", gen_params.cache_params.spectrum_window_size},
-            {"spectrum_flex_window", gen_params.cache_params.spectrum_flex_window},
-            {"spectrum_warmup_steps", gen_params.cache_params.spectrum_warmup_steps},
-            {"spectrum_stop_percent", gen_params.cache_params.spectrum_stop_percent},
-        };
-    }
-
-    if (gen_params.vae_tiling_params.enabled) {
-        root["vae_tiling"] = {
-            {"enabled", gen_params.vae_tiling_params.enabled},
-            {"tile_size_x", gen_params.vae_tiling_params.tile_size_x},
-            {"tile_size_y", gen_params.vae_tiling_params.tile_size_y},
-            {"target_overlap", gen_params.vae_tiling_params.target_overlap},
-            {"rel_size_x", gen_params.vae_tiling_params.rel_size_x},
-            {"rel_size_y", gen_params.vae_tiling_params.rel_size_y},
-        };
-    }
-
-    return root.dump();
-}
-
-std::string get_image_params(const SDContextParams& ctx_params,
-                             const SDGenerationParams& gen_params,
-                             int64_t seed,
-                             SDMode mode) {
     std::string parameter_string;
     if (gen_params.prompt_with_lora.size() != 0) {
         parameter_string += gen_params.prompt_with_lora + "\n";
@@ -2500,7 +2117,7 @@ std::string get_image_params(const SDContextParams& ctx_params,
     parameter_string += "Steps: " + std::to_string(gen_params.sample_params.sample_steps) + ", ";
     parameter_string += "CFG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
     if (gen_params.sample_params.guidance.slg.scale != 0 && gen_params.skip_layers.size() != 0) {
-        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.slg.scale) + ", ";
+        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
         parameter_string += "Skip layers: [";
         for (const auto& layer : gen_params.skip_layers) {
             parameter_string += std::to_string(layer) + ", ";
@@ -2511,9 +2128,6 @@ std::string get_image_params(const SDContextParams& ctx_params,
     }
     parameter_string += "Guidance: " + std::to_string(gen_params.sample_params.guidance.distilled_guidance) + ", ";
     parameter_string += "Eta: " + std::to_string(gen_params.sample_params.eta) + ", ";
-    if (!gen_params.extra_sample_args.empty()) {
-        parameter_string += "Extra sample args: " + gen_params.extra_sample_args + ", ";
-    }
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(gen_params.get_resolved_width()) + "x" + std::to_string(gen_params.get_resolved_height()) + ", ";
     parameter_string += "Model: " + sd_basename(ctx_params.model_path) + ", ";
@@ -2548,14 +2162,6 @@ std::string get_image_params(const SDContextParams& ctx_params,
     if (gen_params.clip_skip != -1) {
         parameter_string += "Clip skip: " + std::to_string(gen_params.clip_skip) + ", ";
     }
-    if (gen_params.hires_enabled) {
-        parameter_string += "Hires upscale: " + gen_params.hires_upscaler + ", ";
-        parameter_string += "Hires scale: " + std::to_string(gen_params.hires_scale) + ", ";
-        parameter_string += "Hires resize: " + std::to_string(gen_params.hires_width) + "x" + std::to_string(gen_params.hires_height) + ", ";
-        parameter_string += "Hires steps: " + std::to_string(gen_params.hires_steps) + ", ";
-        parameter_string += "Denoising strength: " + std::to_string(gen_params.hires_denoising_strength) + ", ";
-    }
     parameter_string += "Version: stable-diffusion.cpp";
-    parameter_string += ", SDCPP: " + build_sdcpp_image_metadata_json(ctx_params, gen_params, seed, mode);
     return parameter_string;
 }

examples/common/common.h

@@ -101,7 +101,6 @@ struct SDContextParams {
     sd_type_t wtype = SD_TYPE_COUNT;
     std::string tensor_type_rules;
     std::string lora_model_dir = ".";
-    std::string hires_upscalers_dir;
 
     std::map<std::string, std::string> embedding_map;
     std::vector<sd_embedding_t> embedding_vec;
@@ -109,7 +108,6 @@ struct SDContextParams {
     rng_type_t rng_type         = CUDA_RNG;
     rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
     bool offload_params_to_cpu  = false;
-    float max_vram              = 0.f;
     bool enable_mmap            = false;
     bool control_net_cpu        = false;
     bool clip_on_cpu            = false;
@@ -168,8 +166,6 @@ struct SDGenerationParams {
 
     sd_sample_params_t sample_params;
     sd_sample_params_t high_noise_sample_params;
-    std::string extra_sample_args;
-    std::string high_noise_extra_sample_args;
     std::vector<int> skip_layers            = {7, 8, 9};
     std::vector<int> high_noise_skip_layers = {7, 8, 9};
 
@@ -194,23 +190,12 @@ struct SDGenerationParams {
     int upscale_repeats   = 1;
     int upscale_tile_size = 128;
 
-    bool hires_enabled         = false;
-    std::string hires_upscaler = "Latent";
-    std::string hires_upscaler_model_path;
-    float hires_scale              = 2.f;
-    int hires_width                = 0;
-    int hires_height               = 0;
-    int hires_steps                = 0;
-    float hires_denoising_strength = 0.7f;
-    int hires_upscale_tile_size    = 128;
-
     std::map<std::string, float> lora_map;
     std::map<std::string, float> high_noise_lora_map;
 
     // Derived and normalized fields.
     std::string prompt_with_lora;  // for metadata record only
     std::vector<sd_lora_t> lora_vec;
-    sd_hires_upscaler_t resolved_hires_upscaler;
 
     // Owned execution payload.
     SDImageOwner init_image;
@@ -240,25 +225,15 @@ struct SDGenerationParams {
     void set_width_and_height_if_unset(int w, int h);
     int get_resolved_width() const;
     int get_resolved_height() const;
-    bool resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict = false);
+    bool resolve(const std::string& lora_model_dir, bool strict = false);
     bool validate(SDMode mode);
-    bool resolve_and_validate(SDMode mode,
+    bool resolve_and_validate(SDMode mode, const std::string& lora_model_dir, bool strict = false);
-                              const std::string& lora_model_dir,
-                              const std::string& hires_upscalers_dir,
-                              bool strict = false);
     sd_img_gen_params_t to_sd_img_gen_params_t();
     sd_vid_gen_params_t to_sd_vid_gen_params_t();
     std::string to_string() const;
 };
 
 std::string version_string();
-std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
+std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed);
-                                            const SDGenerationParams& gen_params,
-                                            int64_t seed,
-                                            SDMode mode = IMG_GEN);
-std::string get_image_params(const SDContextParams& ctx_params,
-                             const SDGenerationParams& gen_params,
-                             int64_t seed,
-                             SDMode mode = IMG_GEN);
 
 #endif  // __EXAMPLES_COMMON_COMMON_H__

examples/server/README.md

@@ -136,8 +136,7 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
-                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -149,18 +148,16 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
-  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           then threads will be set to the number of CPU physical cores
+                                           CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-                                           graph splitting
+  --vae-tiling                             process vae in tiles to reduce memory usage
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
-  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
-                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -175,19 +172,20 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           q4_K). If not specified, the default is the type of the weight file
+                                           type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           immediately will be used.The immediately mode may have precision and
-                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           immediately mode may have precision and compatibility issues with quantized
+                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           other hand, is exactly the opposite.
-                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-                                           opposite.
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
+                                           (overrides --vae-tile-size)
 
 Default Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -196,99 +194,65 @@ Default Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           inside should be stored as images in lexicographical (character) order. For
+                                           lexicographical (character) order. For example, if the control video path is
-                                           example, if the control video path is `frames`, the directory contain images
+                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
-                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
-  --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
-                                           (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
-                                           antialiased), or a model name under --hires-upscalers-dir (default: Latent)
-  --extra-sample-args <string>             extra sampler args, key=value list. Currently lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
+                                           will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
+                                           NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
-  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
-  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
-  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
-  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
-                                           128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
-                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           disabled, a value of 2.5 is nice for sd3.5 medium
+                                           medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
-                                           res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
-                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
-                                           (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
-                                           0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
-                                           res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
-  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
-                                           destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
-                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
-  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
-  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
-  --increase-ref-index                     automatically increase the indices of references images based on the order
-                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
-  --vae-tiling                             process vae in tiles to reduce memory usage
-  --hires                                  enable highres fix
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
-  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
+                                           tcd, res_multistep, res_2s, er_sde] (default: euler for Flux/SD3/Wan, euler_a
-                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           otherwise)
-  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
+                                           ddim_trailing, tcd, res_multistep, res_2s, er_sde] default: euler for Flux/SD3/Wan,
-                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           euler_a otherwise
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
+                                           kl_optimal, lcm, bong_tangent], default: discrete
-                                           discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
-                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
-                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
-                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
-                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           "threshold=0.25" or "threshold=1.5,reset=0"
-                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
-                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
-                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```

examples/server/api.md

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

examples/server/main.cpp

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

examples/server/routes_openai.cpp

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

examples/server/routes_sdapi.cpp

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

examples/server/routes_sdcpp.cpp

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

examples/server/runtime.cpp

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

examples/server/runtime.h

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

include/stable-diffusion.h

@@ -51,8 +51,6 @@ enum sample_method_t {
     RES_MULTISTEP_SAMPLE_METHOD,
     RES_2S_SAMPLE_METHOD,
     ER_SDE_SAMPLE_METHOD,
-    EULER_CFG_PP_SAMPLE_METHOD,
-    EULER_A_CFG_PP_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -205,7 +203,6 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
-    float max_vram;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -240,7 +237,6 @@ typedef struct {
     float* custom_sigmas;
     int custom_sigmas_count;
     float flow_shift;
-    const char* extra_sample_args;
 } sd_sample_params_t;
 
 typedef struct {
@@ -293,32 +289,6 @@ typedef struct {
     const char* path;
 } sd_lora_t;
 
-enum sd_hires_upscaler_t {
-    SD_HIRES_UPSCALER_NONE,
-    SD_HIRES_UPSCALER_LATENT,
-    SD_HIRES_UPSCALER_LATENT_NEAREST,
-    SD_HIRES_UPSCALER_LATENT_NEAREST_EXACT,
-    SD_HIRES_UPSCALER_LATENT_ANTIALIASED,
-    SD_HIRES_UPSCALER_LATENT_BICUBIC,
-    SD_HIRES_UPSCALER_LATENT_BICUBIC_ANTIALIASED,
-    SD_HIRES_UPSCALER_LANCZOS,
-    SD_HIRES_UPSCALER_NEAREST,
-    SD_HIRES_UPSCALER_MODEL,
-    SD_HIRES_UPSCALER_COUNT,
-};
-
-typedef struct {
-    bool enabled;
-    enum sd_hires_upscaler_t upscaler;
-    const char* model_path;
-    float scale;
-    int target_width;
-    int target_height;
-    int steps;
-    float denoising_strength;
-    int upscale_tile_size;
-} sd_hires_params_t;
-
 typedef struct {
     const sd_lora_t* loras;
     uint32_t lora_count;
@@ -342,7 +312,6 @@ typedef struct {
     sd_pm_params_t pm_params;
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
-    sd_hires_params_t hires;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -396,11 +365,8 @@ SD_API const char* sd_preview_name(enum preview_t preview);
 SD_API enum preview_t str_to_preview(const char* str);
 SD_API const char* sd_lora_apply_mode_name(enum lora_apply_mode_t mode);
 SD_API enum lora_apply_mode_t str_to_lora_apply_mode(const char* str);
-SD_API const char* sd_hires_upscaler_name(enum sd_hires_upscaler_t upscaler);
-SD_API enum sd_hires_upscaler_t str_to_sd_hires_upscaler(const char* str);
 
 SD_API void sd_cache_params_init(sd_cache_params_t* cache_params);
-SD_API void sd_hires_params_init(sd_hires_params_t* hires_params);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);

src/anima.hpp

@@ -499,15 +499,9 @@ namespace Anima {
                 encoder_hidden_states = adapted_context;
             }
 
-            sd::ggml_graph_cut::mark_graph_cut(x, "anima.prelude", "x");
-            sd::ggml_graph_cut::mark_graph_cut(embedded_timestep, "anima.prelude", "embedded_timestep");
-            sd::ggml_graph_cut::mark_graph_cut(temb, "anima.prelude", "temb");
-            sd::ggml_graph_cut::mark_graph_cut(encoder_hidden_states, "anima.prelude", "context");
-
             for (int i = 0; i < num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["blocks." + std::to_string(i)]);
                 x          = block->forward(ctx, x, encoder_hidden_states, embedded_timestep, temb, image_pe);
-                sd::ggml_graph_cut::mark_graph_cut(x, "anima.blocks." + std::to_string(i), "x");
             }
 
             x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]

src/auto_encoder_kl.hpp

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

src/clip.hpp

@@ -96,8 +96,7 @@ public:
     ggml_tensor* forward(GGMLRunnerContext* ctx,
                          ggml_tensor* x,
                          ggml_tensor* mask = nullptr,
-                         int clip_skip                       = -1,
+                         int clip_skip     = -1) {
-                         const std::string& graph_cut_prefix = "") {
         // x: [N, n_token, d_model]
         int layer_idx = n_layer - 1;
         // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -113,9 +112,6 @@ public:
             std::string name = "layers." + std::to_string(i);
             auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
             x                = layer->forward(ctx, x, mask);  // [N, n_token, d_model]
-            if (!graph_cut_prefix.empty()) {
-                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".layers." + std::to_string(i), "x");
-            }
             // LOG_DEBUG("layer %d", i);
         }
         return x;
@@ -308,8 +304,7 @@ public:
         auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
 
         auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
-        sd::ggml_graph_cut::mark_graph_cut(x, "clip_text.prelude", "x");
+        x      = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip);
-        x = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip, "clip_text");
         if (return_pooled || with_final_ln) {
             x = final_layer_norm->forward(ctx, x);
         }
@@ -373,8 +368,7 @@ public:
 
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
-        sd::ggml_graph_cut::mark_graph_cut(x, "clip_vision.prelude", "x");
+        x      = encoder->forward(ctx, x, nullptr, clip_skip);
-        x = encoder->forward(ctx, x, nullptr, clip_skip, "clip_vision");
 
         auto last_hidden_state = x;
 

src/common_block.hpp

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

src/conditioner.hpp

@@ -14,12 +14,6 @@ struct SDCondition {
     sd::Tensor<float> c_concat;
     sd::Tensor<int32_t> c_t5_ids;
     sd::Tensor<float> c_t5_weights;
-    sd::Tensor<int32_t> c_input_ids;
-    sd::Tensor<int32_t> c_position_ids;
-    sd::Tensor<int32_t> c_token_types;
-    sd::Tensor<int32_t> c_vinput_mask;
-    std::vector<std::pair<int, sd::Tensor<float>>> c_image_embeds;
-    std::vector<sd::Tensor<float>> c_ref_images;
 
     std::vector<sd::Tensor<float>> extra_c_crossattns;
 
@@ -32,24 +26,10 @@ struct SDCondition {
 
     bool empty() const {
         if (!c_crossattn.empty() || !c_vector.empty() || !c_concat.empty() ||
-            !c_t5_ids.empty() || !c_t5_weights.empty() ||
+            !c_t5_ids.empty() || !c_t5_weights.empty()) {
-            !c_input_ids.empty() || !c_position_ids.empty() ||
-            !c_token_types.empty() || !c_vinput_mask.empty()) {
             return false;
         }
 
-        for (const auto& image_embed : c_image_embeds) {
-            if (!image_embed.second.empty()) {
-                return false;
-            }
-        }
-
-        for (const auto& tensor : c_ref_images) {
-            if (!tensor.empty()) {
-                return false;
-            }
-        }
-
         for (const auto& tensor : extra_c_crossattns) {
             if (!tensor.empty()) {
                 return false;
@@ -105,7 +85,6 @@ public:
     virtual void free_params_buffer()                                                      = 0;
     virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)           = 0;
     virtual size_t get_params_buffer_size()                                                = 0;
-    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
     virtual void set_flash_attention_enabled(bool enabled)                                 = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(int n_threads,
@@ -186,13 +165,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return buffer_size;
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        text_model->set_max_graph_vram_bytes(max_vram_bytes);
-        if (sd_version_is_sdxl(version)) {
-            text_model2->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-    }
-
     void set_flash_attention_enabled(bool enabled) override {
         text_model->set_flash_attention_enabled(enabled);
         if (sd_version_is_sdxl(version)) {
@@ -809,18 +781,6 @@ struct SD3CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        if (clip_l) {
-            clip_l->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-        if (clip_g) {
-            clip_g->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-        if (t5) {
-            t5->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-    }
-
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1164,15 +1124,6 @@ struct FluxCLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        if (clip_l) {
-            clip_l->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-        if (t5) {
-            t5->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-    }
-
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1398,12 +1349,6 @@ struct T5CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        if (t5) {
-            t5->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-    }
-
     void set_flash_attention_enabled(bool enabled) override {
         if (t5) {
             t5->set_flash_attention_enabled(enabled);
@@ -1580,10 +1525,6 @@ struct AnimaConditioner : public Conditioner {
         return llm->get_params_buffer_size();
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        llm->set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -1716,10 +1657,6 @@ struct LLMEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        llm->set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }

src/convert.cpp

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

src/denoiser.hpp

@@ -2,7 +2,6 @@
 #define __DENOISER_HPP__
 
 #include <cmath>
-#include <string>
 #include <utility>
 
 #include "ggml_extend.hpp"
@@ -753,7 +752,7 @@ struct Flux2FlowDenoiser : public FluxFlowDenoiser {
     }
 };
 
-typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int, sd::Tensor<float>*)> denoise_cb_t;
+typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int)> denoise_cb_t;
 
 static std::pair<float, float> get_ancestral_step(float sigma_from,
                                                   float sigma_to,
@@ -809,18 +808,6 @@ static std::tuple<float, float, float> get_ancestral_step_flow(float sigma_from,
     return {sigma_down, sigma_up, alpha_scale};
 }
 
-static std::tuple<float, float, float> get_ancestral_step(float sigma_from,
-                                                          float sigma_to,
-                                                          float eta,
-                                                          bool is_flow_denoiser) {
-    if (is_flow_denoiser) {
-        return get_ancestral_step_flow(sigma_from, sigma_to, eta);
-    } else {
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
-        return {sigma_down, sigma_up, 1.0f};
-    }
-}
-
 static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
                                                 sd::Tensor<float> x,
                                                 const std::vector<float>& sigmas,
@@ -829,7 +816,7 @@ static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        auto denoised_opt = model(x, sigma, i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -852,7 +839,7 @@ static sd::Tensor<float> sample_euler_flow(denoise_cb_t model,
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        auto denoised_opt = model(x, sigma, i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -874,7 +861,7 @@ static sd::Tensor<float> sample_euler(denoise_cb_t model,
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        auto denoised_opt = model(x, sigma, i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -890,7 +877,7 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
                                      const std::vector<float>& sigmas) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
+        auto denoised_opt = model(x, sigmas[i], -(i + 1));
         if (denoised_opt.empty()) {
             return {};
         }
@@ -901,7 +888,7 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
             x += d * dt;
         } else {
             sd::Tensor<float> x2 = x + d * dt;
-            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1, nullptr);
+            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -918,7 +905,7 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
                                      const std::vector<float>& sigmas) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
+        auto denoised_opt = model(x, sigmas[i], -(i + 1));
         if (denoised_opt.empty()) {
             return {};
         }
@@ -931,7 +918,7 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
             float dt_1           = sigma_mid - sigmas[i];
             float dt_2           = sigmas[i + 1] - sigmas[i];
             sd::Tensor<float> x2 = x + d * dt_1;
-            auto denoised2_opt   = model(x2, sigma_mid, i + 1, nullptr);
+            auto denoised2_opt   = model(x2, sigma_mid, i + 1);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -952,7 +939,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
+        auto denoised_opt = model(x, sigmas[i], -(i + 1));
         if (denoised_opt.empty()) {
             return {};
         }
@@ -968,7 +955,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
             float s              = t + 0.5f * h;
             float sigma_s        = sigma_fn(s);
             sd::Tensor<float> x2 = (sigma_s / sigma_fn(t)) * x - (exp(-h * 0.5f) - 1) * denoised;
-            auto denoised2_opt   = model(x2, sigma_s, i + 1, nullptr);
+            auto denoised2_opt   = model(x2, sigma_s, i + 1);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -995,7 +982,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
 
         bool opt_first_step = (1.0 - sigma < 1e-6);
 
-        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1), nullptr);
+        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1));
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1024,8 +1011,8 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                 // so sigma_s = 1 = sigma, and sigma_s_i_ratio = sigma_s / sigma = 1
                 // u = (x*sigma_s_i_ratio)+(denoised*(1.0f-sigma_s_i_ratio))
                 //   = (x*1)+(denoised*0) = x
-                // so D_i = model(u, sigma_s, i + 1, nullptr)
+                // so D_i = model(u, sigma_s, i + 1)
-                //        = model(x, sigma,   i + 1, nullptr)
+                //        = model(x, sigma,   i + 1)
                 //        = denoised
                 D_i = denoised;
 
@@ -1058,7 +1045,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                 float sigma_s_i_ratio = sigma_s / sigma;
                 sd::Tensor<float> u   = (x * sigma_s_i_ratio) + (denoised * (1.0f - sigma_s_i_ratio));
 
-                auto denoised2_opt = model(u, sigma_s, i + 1, nullptr);
+                auto denoised2_opt = model(u, sigma_s, i + 1);
                 if (denoised2_opt.empty()) {
                     return {};
                 }
@@ -1085,7 +1072,7 @@ static sd::Tensor<float> sample_dpmpp_2m(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
+        auto denoised_opt = model(x, sigmas[i], i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1117,7 +1104,7 @@ static sd::Tensor<float> sample_dpmpp_2m_v2(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
+        auto denoised_opt = model(x, sigmas[i], i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1149,83 +1136,10 @@ static sd::Tensor<float> sample_lcm(denoise_cb_t model,
                                     sd::Tensor<float> x,
                                     const std::vector<float>& sigmas,
                                     std::shared_ptr<RNG> rng,
-                                    bool is_flow_denoiser,
+                                    bool is_flow_denoiser) {
-                                    const char* extra_sample_args = nullptr) {
-    struct LCMSampleArgs {
-        float noise_clip_std    = 0.0f;
-        float noise_scale_start = 1.0f;
-        float noise_scale_end   = 1.0f;
-    };
-
-    auto trim = [](std::string value) -> std::string {
-        const char* whitespace = " \t\r\n";
-        size_t begin           = value.find_first_not_of(whitespace);
-        if (begin == std::string::npos) {
-            return "";
-        }
-        size_t end = value.find_last_not_of(whitespace);
-        return value.substr(begin, end - begin + 1);
-    };
-
-    LCMSampleArgs args;
-    if (extra_sample_args != nullptr && extra_sample_args[0] != '\0') {
-        std::string raw(extra_sample_args);
-        size_t start                   = 0;
-        bool noise_scale_end_was_set   = false;
-        bool noise_scale_start_was_set = false;
-        auto parse_arg                 = [&](const std::string& item) {
-            std::string token = trim(item);
-            if (token.empty()) {
-                return;
-            }
-            size_t eq = token.find('=');
-            if (eq == std::string::npos) {
-                LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
-                return;
-            }
-
-            std::string key   = trim(token.substr(0, eq));
-            std::string value = trim(token.substr(eq + 1));
-            float parsed      = 0.0f;
-            try {
-                size_t consumed = 0;
-                parsed          = std::stof(value, &consumed);
-                if (trim(value.substr(consumed)).size() != 0) {
-                    LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
-                    return;
-                }
-            } catch (const std::exception&) {
-                LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
-                return;
-            }
-
-            if (key == "noise_clip_std") {
-                args.noise_clip_std = parsed;
-            } else if (key == "noise_scale_start") {
-                args.noise_scale_start    = parsed;
-                noise_scale_start_was_set = true;
-            } else if (key == "noise_scale_end") {
-                args.noise_scale_end    = parsed;
-                noise_scale_end_was_set = true;
-            } else {
-                LOG_WARN("ignoring unknown lcm extra sample arg '%s'", key.c_str());
-            }
-        };
-
-        for (size_t pos = 0; pos <= raw.size(); ++pos) {
-            if (pos == raw.size() || raw[pos] == ',' || raw[pos] == ';') {
-                parse_arg(raw.substr(start, pos - start));
-                start = pos + 1;
-            }
-        }
-        if (noise_scale_start_was_set && !noise_scale_end_was_set) {
-            args.noise_scale_end = args.noise_scale_start;
-        }
-    }
-
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
+        auto denoised_opt = model(x, sigmas[i], i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1234,27 +1148,7 @@ static sd::Tensor<float> sample_lcm(denoise_cb_t model,
             if (is_flow_denoiser) {
                 x *= (1 - sigmas[i + 1]);
             }
-            auto noise = sd::Tensor<float>::randn_like(x, rng);
+            x += sd::Tensor<float>::randn_like(x, rng) * sigmas[i + 1];
-            if (args.noise_clip_std > 0.0f && noise.numel() > 0) {
-                double mean = 0.0;
-                for (int64_t j = 0; j < noise.numel(); ++j) {
-                    mean += static_cast<double>(noise[j]);
-                }
-                mean /= static_cast<double>(noise.numel());
-
-                double variance = 0.0;
-                for (int64_t j = 0; j < noise.numel(); ++j) {
-                    double centered = static_cast<double>(noise[j]) - mean;
-                    variance += centered * centered;
-                }
-                variance /= static_cast<double>(noise.numel());
-
-                float clip_val = args.noise_clip_std * static_cast<float>(std::sqrt(variance));
-                noise          = sd::ops::clamp(noise, -clip_val, clip_val);
-            }
-            float t           = steps > 1 ? static_cast<float>(i) / static_cast<float>(steps - 1) : 0.0f;
-            float noise_scale = args.noise_scale_start + (args.noise_scale_end - args.noise_scale_start) * t;
-            x += noise * (sigmas[i + 1] * noise_scale);
         }
     }
     return x;
@@ -1271,7 +1165,7 @@ static sd::Tensor<float> sample_ipndm(denoise_cb_t model,
         float sigma      = sigmas[i];
         float sigma_next = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        auto denoised_opt = model(x, sigma, i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1315,7 +1209,7 @@ static sd::Tensor<float> sample_ipndm_v(denoise_cb_t model,
         float sigma  = sigmas[i];
         float t_next = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        auto denoised_opt = model(x, sigma, i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1353,7 +1247,6 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
                                               sd::Tensor<float> x,
                                               const std::vector<float>& sigmas,
                                               std::shared_ptr<RNG> rng,
-                                              bool is_flow_denoiser,
                                               float eta) {
     sd::Tensor<float> old_denoised = x;
     bool have_old_sigma            = false;
@@ -1377,7 +1270,7 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
+        auto denoised_opt = model(x, sigmas[i], i + 1);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1385,8 +1278,7 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
 
         float sigma_from            = sigmas[i];
         float sigma_to              = sigmas[i + 1];
-
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
-        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);
 
         if (sigma_down == 0.0f || !have_old_sigma) {
             x += ((x - denoised) / sigma_from) * (sigma_down - sigma_from);
@@ -1413,10 +1305,7 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
             x = sigma_fn(h) * x + h * (b1 * denoised + b2 * old_denoised);
         }
 
-        if (sigma_to > 0.0f && sigma_up > 0.0f) {
+        if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
-            if (is_flow_denoiser) {
-                x *= alpha_scale;
-            }
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
 
@@ -1431,7 +1320,6 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
                                        sd::Tensor<float> x,
                                        const std::vector<float>& sigmas,
                                        std::shared_ptr<RNG> rng,
-                                       bool is_flow_denoiser,
                                        float eta) {
     const float c2 = 0.5f;
     auto t_fn      = [](float sigma) -> float { return -logf(sigma); };
@@ -1454,13 +1342,13 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
         float sigma_from = sigmas[i];
         float sigma_to   = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma_from, -(i + 1), nullptr);
+        auto denoised_opt = model(x, sigma_from, -(i + 1));
         if (denoised_opt.empty()) {
             return {};
         }
         sd::Tensor<float> denoised = std::move(denoised_opt);
 
-        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
 
         sd::Tensor<float> x0 = x;
         if (sigma_down == 0.0f || sigma_from == 0.0f) {
@@ -1480,7 +1368,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
             sd::Tensor<float> eps1 = denoised - x0;
             sd::Tensor<float> x2   = x0 + eps1 * (h * a21);
 
-            auto denoised2_opt = model(x2, sigma_c2, i + 1, nullptr);
+            auto denoised2_opt = model(x2, sigma_c2, i + 1);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -1489,10 +1377,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
             x                           = x0 + h * (b1 * eps1 + b2 * eps2);
         }
 
-        if (sigma_to > 0.0f && sigma_up > 0.0f) {
+        if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
-            if (is_flow_denoiser) {
-                x *= alpha_scale;
-            }
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
     }
@@ -1557,7 +1442,7 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1, nullptr);
+        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1);
         if (denoised.empty()) {
             return {};
         }
@@ -1643,7 +1528,7 @@ static sd::Tensor<float> sample_ddim_trailing(denoise_cb_t model,
         float sigma    = sigmas[i];
         float sigma_to = sigmas[i + 1];
 
-        auto model_output_opt = model(x, sigma, i + 1, nullptr);
+        auto model_output_opt = model(x, sigma, i + 1);
         if (model_output_opt.empty()) {
             return {};
         }
@@ -1715,7 +1600,7 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
         int timestep_s    = (int)floor((1 - eta) * prev_timestep);
         float sigma       = sigmas[i];
 
-        auto model_output_opt = model(x, sigma, i + 1, nullptr);
+        auto model_output_opt = model(x, sigma, i + 1);
         if (model_output_opt.empty()) {
             return {};
         }
@@ -1743,56 +1628,6 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
     return x;
 }
 
-static sd::Tensor<float> sample_euler_cfg_pp(denoise_cb_t model,
-                                             sd::Tensor<float> x,
-                                             const std::vector<float>& sigmas) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma = sigmas[i];
-        sd::Tensor<float> uncond_denoised;
-
-        auto denoised_opt = model(x, sigma, i + 1, &uncond_denoised);
-        if (denoised_opt.empty() || uncond_denoised.empty()) {
-            return {};
-        }
-
-        sd::Tensor<float> denoised = std::move(denoised_opt);
-        sd::Tensor<float> d        = (x - uncond_denoised) / sigma;
-
-        x = denoised + d * sigmas[i + 1];
-    }
-    return x;
-}
-
-static sd::Tensor<float> sample_euler_ancestral_cfg_pp(denoise_cb_t model,
-                                                       sd::Tensor<float> x,
-                                                       const std::vector<float>& sigmas,
-                                                       std::shared_ptr<RNG> rng,
-                                                       float eta) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma = sigmas[i];
-        sd::Tensor<float> uncond_denoised;
-
-        auto denoised_opt = model(x, sigma, i + 1, &uncond_denoised);
-        if (denoised_opt.empty() || uncond_denoised.empty()) {
-            return {};
-        }
-
-        sd::Tensor<float> denoised = std::move(denoised_opt);
-        sd::Tensor<float> d        = (x - uncond_denoised) / sigma;
-
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
-
-        x = denoised + d * sigma_down;
-
-        if (sigmas[i + 1] > 0) {
-            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
-        }
-    }
-    return x;
-}
-
 // k diffusion reverse ODE: dx = (x - D(x;\sigma)) / \sigma dt; \sigma(t) = t
 static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                             denoise_cb_t model,
@@ -1800,8 +1635,7 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                             std::vector<float> sigmas,
                                             std::shared_ptr<RNG> rng,
                                             float eta,
-                                            bool is_flow_denoiser,
+                                            bool is_flow_denoiser) {
-                                            const char* extra_sample_args) {
     switch (method) {
         case EULER_A_SAMPLE_METHOD:
             if (is_flow_denoiser)
@@ -1824,25 +1658,21 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case DPMPP2Mv2_SAMPLE_METHOD:
             return sample_dpmpp_2m_v2(model, std::move(x), sigmas);
         case LCM_SAMPLE_METHOD:
-            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser, extra_sample_args);
+            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser);
         case IPNDM_SAMPLE_METHOD:
             return sample_ipndm(model, std::move(x), sigmas);
         case IPNDM_V_SAMPLE_METHOD:
             return sample_ipndm_v(model, std::move(x), sigmas);
         case RES_MULTISTEP_SAMPLE_METHOD:
-            return sample_res_multistep(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
+            return sample_res_multistep(model, std::move(x), sigmas, rng, eta);
         case RES_2S_SAMPLE_METHOD:
-            return sample_res_2s(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
+            return sample_res_2s(model, std::move(x), sigmas, rng, eta);
         case ER_SDE_SAMPLE_METHOD:
             return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case DDIM_TRAILING_SAMPLE_METHOD:
             return sample_ddim_trailing(model, std::move(x), sigmas, rng, eta);
         case TCD_SAMPLE_METHOD:
             return sample_tcd(model, std::move(x), sigmas, rng, eta);
-        case EULER_CFG_PP_SAMPLE_METHOD:
-            return sample_euler_cfg_pp(model, std::move(x), sigmas);
-        case EULER_A_CFG_PP_SAMPLE_METHOD:
-            return sample_euler_ancestral_cfg_pp(model, std::move(x), sigmas, rng, eta);
         default:
             return {};
     }

src/diffusion_model.hpp

@@ -5,7 +5,6 @@
 #include "anima.hpp"
 #include "ernie_image.hpp"
 #include "flux.hpp"
-#include "hidream_o1.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
 #include "tensor_ggml.hpp"
@@ -23,12 +22,6 @@ struct DiffusionParams {
     const sd::Tensor<float>* t5_weights               = nullptr;
     const sd::Tensor<float>* guidance                 = nullptr;
     const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
-    const sd::Tensor<int32_t>* input_ids                               = nullptr;
-    const sd::Tensor<int32_t>* input_pos                               = nullptr;
-    const sd::Tensor<int32_t>* token_types                             = nullptr;
-    const sd::Tensor<int32_t>* vinput_mask                             = nullptr;
-    const std::vector<sd::Tensor<float>>* vlm_images                   = nullptr;
-    const std::vector<std::pair<int, sd::Tensor<float>>>* image_embeds = nullptr;
     bool increase_ref_index                           = false;
     int num_video_frames                              = -1;
     const std::vector<sd::Tensor<float>>* controls    = nullptr;
@@ -56,7 +49,6 @@ struct DiffusionModel {
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
     virtual int64_t get_adm_in_channels()                            = 0;
     virtual void set_flash_attention_enabled(bool enabled)           = 0;
-    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes)     = 0;
     virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };
 
@@ -106,10 +98,6 @@ struct UNetModel : public DiffusionModel {
         unet.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        unet.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         unet.set_circular_axes(circular_x, circular_y);
     }
@@ -176,10 +164,6 @@ struct MMDiTModel : public DiffusionModel {
         mmdit.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        mmdit.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         mmdit.set_circular_axes(circular_x, circular_y);
     }
@@ -245,10 +229,6 @@ struct FluxModel : public DiffusionModel {
         flux.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        flux.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         flux.set_circular_axes(circular_x, circular_y);
     }
@@ -319,10 +299,6 @@ struct AnimaModel : public DiffusionModel {
         anima.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        anima.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         anima.set_circular_axes(circular_x, circular_y);
     }
@@ -388,10 +364,6 @@ struct WanModel : public DiffusionModel {
         wan.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        wan.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         wan.set_circular_axes(circular_x, circular_y);
     }
@@ -461,10 +433,6 @@ struct QwenImageModel : public DiffusionModel {
         qwen_image.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        qwen_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         qwen_image.set_circular_axes(circular_x, circular_y);
     }
@@ -483,82 +451,6 @@ struct QwenImageModel : public DiffusionModel {
     }
 };
 
-struct HiDreamO1Model : public DiffusionModel {
-    std::string prefix;
-    HiDreamO1::HiDreamO1Runner hidream_o1;
-
-    HiDreamO1Model(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
-                   const String2TensorStorage& tensor_storage_map = {},
-                   const std::string& prefix                      = "model")
-        : prefix(prefix), hidream_o1(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return hidream_o1.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        hidream_o1.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        hidream_o1.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        hidream_o1.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        hidream_o1.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return hidream_o1.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        hidream_o1.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 0;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        hidream_o1.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        hidream_o1.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        hidream_o1.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        GGML_ASSERT(diffusion_params.input_ids != nullptr);
-        GGML_ASSERT(diffusion_params.input_pos != nullptr);
-        GGML_ASSERT(diffusion_params.token_types != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_images;
-        static const std::vector<std::pair<int, sd::Tensor<float>>> empty_image_embeds;
-        return hidream_o1.compute(n_threads,
-                                  *diffusion_params.x,
-                                  *diffusion_params.timesteps,
-                                  *diffusion_params.input_ids,
-                                  *diffusion_params.input_pos,
-                                  *diffusion_params.token_types,
-                                  tensor_or_empty(diffusion_params.vinput_mask),
-                                  diffusion_params.image_embeds ? *diffusion_params.image_embeds : empty_image_embeds,
-                                  diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_images);
-    }
-};
-
 struct ZImageModel : public DiffusionModel {
     std::string prefix;
     ZImage::ZImageRunner z_image;
@@ -607,10 +499,6 @@ struct ZImageModel : public DiffusionModel {
         z_image.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        z_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         z_image.set_circular_axes(circular_x, circular_y);
     }
@@ -676,10 +564,6 @@ struct ErnieImageModel : public DiffusionModel {
         ernie_image.set_flash_attention_enabled(enabled);
     }
 
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        ernie_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) override {
         ernie_image.set_circular_axes(circular_x, circular_y);
     }

src/ernie_image.hpp

@@ -295,8 +295,6 @@ namespace ErnieImage {
             auto c      = time_embedding->forward(ctx, sample);  // [N, hidden_size]
 
             auto mod_params = adaLN_mod->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, 6 * hidden_size]
-            sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.prelude", "hidden_states");
-            // sd::ggml_graph_cut::mark_graph_cut(mod_params, "ernie_image.prelude", "mod_params");
             auto chunks     = ggml_ext_chunk(ctx->ggml_ctx, mod_params, 6, 0);
             std::vector<ggml_tensor*> temb;
             temb.reserve(6);
@@ -307,7 +305,6 @@ namespace ErnieImage {
             for (int i = 0; i < params.num_layers; i++) {
                 auto layer    = std::dynamic_pointer_cast<ErnieImageSharedAdaLNBlock>(blocks["layers." + std::to_string(i)]);
                 hidden_states = layer->forward(ctx, hidden_states, pe, temb);
-                sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.layers." + std::to_string(i), "hidden_states");
             }
 
             hidden_states = final_norm->forward(ctx, hidden_states, c);

src/esrgan.hpp

@@ -125,32 +125,26 @@ public:
         auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
 
         auto feat      = conv_first->forward(ctx, x);
-        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.prelude", "feat");
         auto body_feat = feat;
         for (int i = 0; i < num_block; i++) {
             std::string name = "body." + std::to_string(i);
             auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
 
             body_feat = block->forward(ctx, body_feat);
-            sd::ggml_graph_cut::mark_graph_cut(body_feat, "esrgan.body." + std::to_string(i), "feat");
         }
         body_feat = conv_body->forward(ctx, body_feat);
         feat      = ggml_add(ctx->ggml_ctx, feat, body_feat);
-        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.body.out", "feat");
         // upsample
         if (scale >= 2) {
             auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
             feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
-            sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up1", "feat");
             if (scale == 4) {
                 auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
                 feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
-                sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up2", "feat");
             }
         }
         // for all scales
         auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
-        sd::ggml_graph_cut::mark_graph_cut(out, "esrgan.final", "out");
         return out;
     }
 };

src/flux.hpp

@@ -928,9 +928,6 @@ namespace Flux {
             }
 
             txt = txt_in->forward(ctx, txt);
-            sd::ggml_graph_cut::mark_graph_cut(img, "flux.prelude", "img");
-            sd::ggml_graph_cut::mark_graph_cut(txt, "flux.prelude", "txt");
-            sd::ggml_graph_cut::mark_graph_cut(vec, "flux.prelude", "vec");
 
             for (int i = 0; i < params.depth; i++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
@@ -942,8 +939,6 @@ namespace Flux {
                 auto img_txt = block->forward(ctx, img, txt, vec, pe, txt_img_mask, ds_img_mods, ds_txt_mods);
                 img          = img_txt.first;   // [N, n_img_token, hidden_size]
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
-                sd::ggml_graph_cut::mark_graph_cut(img, "flux.double_blocks." + std::to_string(i), "img");
-                sd::ggml_graph_cut::mark_graph_cut(txt, "flux.double_blocks." + std::to_string(i), "txt");
             }
 
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
@@ -954,7 +949,6 @@ namespace Flux {
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask, ss_mods);
-                sd::ggml_graph_cut::mark_graph_cut(txt_img, "flux.single_blocks." + std::to_string(i), "txt_img");
             }
 
             img = ggml_view_3d(ctx->ggml_ctx,

src/ggml_extend_backend.hpp

@@ -1,298 +0,0 @@
-#ifndef __GGML_EXTEND_BACKEND_HPP__
-#define __GGML_EXTEND_BACKEND_HPP__
-
-#include <cstring>
-#include <mutex>
-
-#include "ggml-backend.h"
-#include "ggml.h"
-
-#ifndef __STATIC_INLINE__
-#define __STATIC_INLINE__ static inline
-#endif
-
-inline void ggml_backend_load_all_once() {
-    // If the registry already has devices and the CPU backend is present,
-    // assume either static registration or explicit host-side preloading has
-    // completed and avoid rescanning the default paths.
-    if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
-        return;
-    }
-    // In dynamic-backend mode the backend modules are discovered at runtime,
-    // so we must load them before asking for the CPU backend or its proc table.
-    // If the host preloaded only a subset of backends, allow one default-path
-    // scan so missing modules can still be discovered.
-    static std::once_flag once;
-    std::call_once(once, []() {
-        if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
-            return;
-        }
-        ggml_backend_load_all();
-    });
-}
-
-// Do not gate this branch on GGML_CPU or GGML_CPU_ALL_VARIANTS:
-// those are CMake options used to configure ggml itself, but they are not
-// exported as PUBLIC compile definitions to stable-diffusion in backend-DL mode.
-// In practice, this target can reliably see GGML_BACKEND_DL, but not whether
-// the CPU backend was compiled as a loadable module. We therefore use runtime
-// backend discovery instead of compile-time assumptions.
-
-__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_cpu_reg() {
-    ggml_backend_reg_t reg = ggml_backend_reg_by_name("CPU");
-    if (reg != nullptr) {
-        return reg;
-    }
-
-    ggml_backend_load_all_once();
-    return ggml_backend_reg_by_name("CPU");
-}
-
-__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_reg_from_backend(ggml_backend_t backend) {
-    if (backend != nullptr) {
-        ggml_backend_dev_t device = ggml_backend_get_device(backend);
-        if (device != nullptr) {
-            return ggml_backend_dev_backend_reg(device);
-        }
-    }
-
-    return ggml_backend_cpu_reg();
-}
-
-__STATIC_INLINE__ ggml_backend_t ggml_backend_cpu_init() {
-    ggml_backend_t backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-    if (backend != nullptr) {
-        return backend;
-    }
-
-    ggml_backend_load_all_once();
-    return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-}
-
-__STATIC_INLINE__ bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    if (backend == nullptr) {
-        return false;
-    }
-
-    ggml_backend_dev_t device = ggml_backend_get_device(backend);
-    if (device != nullptr) {
-        return ggml_backend_dev_type(device) == GGML_BACKEND_DEVICE_TYPE_CPU;
-    }
-
-    const char* backend_name = ggml_backend_name(backend);
-    return backend_name != nullptr && std::strcmp(backend_name, "CPU") == 0;
-}
-
-__STATIC_INLINE__ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
-    if (reg == nullptr) {
-        return;
-    }
-
-    auto fn = reinterpret_cast<ggml_backend_set_n_threads_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"));
-    if (fn != nullptr) {
-        fn(backend_cpu, n_threads);
-    }
-}
-
-using __ggml_backend_cpu_set_threadpool_t = void (*)(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-
-__STATIC_INLINE__ void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
-    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
-    if (reg == nullptr) {
-        return;
-    }
-
-    auto fn = reinterpret_cast<__ggml_backend_cpu_set_threadpool_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"));
-    if (fn != nullptr) {
-        fn(backend_cpu, threadpool);
-    }
-}
-
-__STATIC_INLINE__ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void* abort_callback_data) {
-    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
-    if (reg == nullptr) {
-        return;
-    }
-
-    auto fn = reinterpret_cast<ggml_backend_set_abort_callback_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"));
-    if (fn != nullptr) {
-        fn(backend_cpu, abort_callback, abort_callback_data);
-    }
-}
-
-__STATIC_INLINE__ ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
-    if (tensor == nullptr) {
-        return nullptr;
-    }
-
-    return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
-}
-
-__STATIC_INLINE__ bool ggml_backend_tensor_is_host_accessible(const struct ggml_tensor* tensor) {
-    if (tensor == nullptr || tensor->data == nullptr) {
-        return false;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_tensor_buffer(tensor);
-    return buffer == nullptr || ggml_backend_buffer_is_host(buffer);
-}
-
-__STATIC_INLINE__ size_t ggml_backend_tensor_offset(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
-    return (size_t)(i0 * tensor->nb[0] + i1 * tensor->nb[1] + i2 * tensor->nb[2] + i3 * tensor->nb[3]);
-}
-
-template <typename T>
-__STATIC_INLINE__ void ggml_backend_tensor_write_scalar(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, T value) {
-    const size_t offset = ggml_backend_tensor_offset(tensor, i0, i1, i2, i3);
-
-    if (ggml_backend_tensor_is_host_accessible(tensor)) {
-        auto* dst = reinterpret_cast<T*>(reinterpret_cast<char*>(tensor->data) + offset);
-        *dst      = value;
-        return;
-    }
-
-    ggml_backend_tensor_set(const_cast<struct ggml_tensor*>(tensor), &value, offset, sizeof(T));
-}
-
-__STATIC_INLINE__ void ggml_set_f32_nd(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float value) {
-    switch (tensor->type) {
-        case GGML_TYPE_I8:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int8_t>(value));
-            break;
-        case GGML_TYPE_I16:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int16_t>(value));
-            break;
-        case GGML_TYPE_I32:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int32_t>(value));
-            break;
-        case GGML_TYPE_F16:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_fp16(value));
-            break;
-        case GGML_TYPE_BF16:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_bf16(value));
-            break;
-        case GGML_TYPE_F32:
-            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, value);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-}
-
-__STATIC_INLINE__ void ggml_set_f32_1d(const struct ggml_tensor* tensor, int i, float value) {
-    if (!ggml_is_contiguous(tensor)) {
-        int64_t id[4] = {0, 0, 0, 0};
-        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
-        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
-        return;
-    }
-
-    switch (tensor->type) {
-        case GGML_TYPE_I8:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int8_t>(value));
-            break;
-        case GGML_TYPE_I16:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int16_t>(value));
-            break;
-        case GGML_TYPE_I32:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int32_t>(value));
-            break;
-        case GGML_TYPE_F16:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_fp16(value));
-            break;
-        case GGML_TYPE_BF16:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_bf16(value));
-            break;
-        case GGML_TYPE_F32:
-            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, value);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-}
-
-__STATIC_INLINE__ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context* ctx, struct ggml_cgraph* cgraph, int n_threads) {
-    (void)ctx;
-
-    // The legacy ggml_graph_compute_with_ctx() symbol lives in ggml-cpu, but
-    // the backend proc table does not expose it in GGML_BACKEND_DL mode.
-    // Recreate the old behavior by initializing the CPU backend explicitly and
-    // executing the graph through the generic backend API.
-    ggml_backend_t backend = ggml_backend_cpu_init();
-    if (backend == nullptr) {
-        return GGML_STATUS_ALLOC_FAILED;
-    }
-
-    ggml_backend_cpu_set_n_threads(backend, n_threads);
-
-    const enum ggml_status status = ggml_backend_graph_compute(backend, cgraph);
-    ggml_backend_free(backend);
-
-    return status;
-}
-
-__STATIC_INLINE__ ggml_tensor* ggml_set_f32(struct ggml_tensor* tensor, float value) {
-    GGML_ASSERT(tensor != nullptr);
-
-    if (ggml_backend_tensor_is_host_accessible(tensor) && ggml_is_contiguous(tensor)) {
-        const int64_t nelements = ggml_nelements(tensor);
-
-        switch (tensor->type) {
-            case GGML_TYPE_I8: {
-                auto* data     = reinterpret_cast<int8_t*>(tensor->data);
-                const int8_t v = static_cast<int8_t>(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_I16: {
-                auto* data      = reinterpret_cast<int16_t*>(tensor->data);
-                const int16_t v = static_cast<int16_t>(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_I32: {
-                auto* data      = reinterpret_cast<int32_t*>(tensor->data);
-                const int32_t v = static_cast<int32_t>(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_F16: {
-                auto* data          = reinterpret_cast<ggml_fp16_t*>(tensor->data);
-                const ggml_fp16_t v = ggml_fp32_to_fp16(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_BF16: {
-                auto* data          = reinterpret_cast<ggml_bf16_t*>(tensor->data);
-                const ggml_bf16_t v = ggml_fp32_to_bf16(value);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = v;
-                }
-            } break;
-            case GGML_TYPE_F32: {
-                auto* data = reinterpret_cast<float*>(tensor->data);
-                for (int64_t i = 0; i < nelements; ++i) {
-                    data[i] = value;
-                }
-            } break;
-            default:
-                GGML_ABORT("fatal error");
-        }
-
-        return tensor;
-    }
-
-    const int64_t nelements = ggml_nelements(tensor);
-    for (int64_t i = 0; i < nelements; ++i) {
-        ggml_set_f32_1d(tensor, static_cast<int>(i), value);
-    }
-
-    return tensor;
-}
-
-#endif

src/ggml_graph_cut.cpp

@@ -1,700 +0,0 @@
-#include "ggml_graph_cut.h"
-
-#include <algorithm>
-#include <cstring>
-#include <map>
-#include <set>
-#include <sstream>
-#include <stack>
-#include <unordered_map>
-
-#include "ggml-alloc.h"
-#include "ggml-backend.h"
-#include "util.h"
-
-#include "../ggml/src/ggml-impl.h"
-
-namespace sd::ggml_graph_cut {
-
-    static std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
-        if (tensor == nullptr) {
-            return "<null>";
-        }
-        if (tensor->name[0] != '\0') {
-            return tensor->name;
-        }
-        return sd_format("<tensor@%p>", (const void*)tensor);
-    }
-
-    static int graph_leaf_index(ggml_cgraph* gf, const ggml_tensor* tensor) {
-        GGML_ASSERT(gf != nullptr);
-        GGML_ASSERT(tensor != nullptr);
-        for (int i = 0; i < gf->n_leafs; ++i) {
-            if (gf->leafs[i] == tensor) {
-                return i;
-            }
-        }
-        return -1;
-    }
-
-    static bool is_params_tensor(const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                                 const ggml_tensor* tensor) {
-        if (tensor == nullptr) {
-            return false;
-        }
-        return params_tensor_set.find(tensor) != params_tensor_set.end();
-    }
-
-    static int graph_node_index_by_name(ggml_cgraph* gf, const char* name) {
-        GGML_ASSERT(gf != nullptr);
-        if (name == nullptr || name[0] == '\0') {
-            return -1;
-        }
-        const int n_nodes = ggml_graph_n_nodes(gf);
-        for (int i = 0; i < n_nodes; ++i) {
-            ggml_tensor* node = ggml_graph_node(gf, i);
-            if (node != nullptr && std::strcmp(node->name, name) == 0) {
-                return i;
-            }
-        }
-        return -1;
-    }
-
-    static Plan::InputShape input_shape(const ggml_tensor* tensor) {
-        Plan::InputShape shape;
-        if (tensor == nullptr) {
-            return shape;
-        }
-        shape.type = tensor->type;
-        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
-            shape.ne[static_cast<size_t>(i)] = tensor->ne[i];
-        }
-        return shape;
-    }
-
-    static size_t graph_cut_segment_vram_bytes(const Segment& segment) {
-        return segment.compute_buffer_size +
-               segment.input_param_bytes +
-               segment.input_previous_cut_bytes +
-               segment.output_bytes;
-    }
-
-    static Segment make_segment_seed(const Plan& plan,
-                                     size_t start_segment_index,
-                                     size_t end_segment_index) {
-        GGML_ASSERT(start_segment_index < plan.segments.size());
-        GGML_ASSERT(end_segment_index < plan.segments.size());
-        GGML_ASSERT(start_segment_index <= end_segment_index);
-
-        Segment seed;
-        const auto& start_segment  = plan.segments[start_segment_index];
-        const auto& target_segment = plan.segments[end_segment_index];
-        std::unordered_set<int> seen_output_node_indices;
-        for (size_t seg_idx = start_segment_index; seg_idx <= end_segment_index; ++seg_idx) {
-            for (int output_node_index : plan.segments[seg_idx].output_node_indices) {
-                if (seen_output_node_indices.insert(output_node_index).second) {
-                    seed.output_node_indices.push_back(output_node_index);
-                }
-            }
-        }
-        if (start_segment_index == end_segment_index) {
-            seed.group_name = target_segment.group_name;
-        } else {
-            seed.group_name = sd_format("%s..%s",
-                                        start_segment.group_name.c_str(),
-                                        target_segment.group_name.c_str());
-        }
-        return seed;
-    }
-
-    static void build_segment(ggml_cgraph* gf,
-                              Plan& plan,
-                              Segment& segment,
-                              const std::unordered_map<const ggml_tensor*, int>& producer_index,
-                              std::unordered_set<int>& available_cut_output_node_indices,
-                              ggml_backend_t backend,
-                              const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                              const char* log_desc) {
-        std::set<int> internal_nodes;
-        std::unordered_set<const ggml_tensor*> input_seen;
-        std::vector<Segment::InputRef> input_refs;
-
-        std::stack<ggml_tensor*> work_stack;
-        for (int output_node_index : segment.output_node_indices) {
-            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
-            if (output != nullptr) {
-                work_stack.push(output);
-            }
-        }
-
-        while (!work_stack.empty()) {
-            ggml_tensor* tensor = work_stack.top();
-            work_stack.pop();
-
-            if (tensor == nullptr) {
-                continue;
-            }
-
-            auto producer_it = producer_index.find(tensor);
-            if (producer_it == producer_index.end()) {
-                if (input_seen.insert(tensor).second) {
-                    Segment::InputRef input_ref;
-                    input_ref.type         = is_params_tensor(params_tensor_set, tensor) ? Segment::INPUT_PARAM : Segment::INPUT_EXTERNAL;
-                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
-                    input_ref.leaf_index   = graph_leaf_index(gf, tensor);
-                    input_refs.push_back(std::move(input_ref));
-                }
-                continue;
-            }
-
-            int node_idx = producer_it->second;
-            if (available_cut_output_node_indices.find(node_idx) != available_cut_output_node_indices.end()) {
-                if (input_seen.insert(tensor).second) {
-                    Segment::InputRef input_ref;
-                    input_ref.type         = Segment::INPUT_PREVIOUS_CUT;
-                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
-                    input_ref.node_index   = node_idx;
-                    input_refs.push_back(std::move(input_ref));
-                }
-                continue;
-            }
-
-            if (!internal_nodes.insert(node_idx).second) {
-                continue;
-            }
-
-            ggml_tensor* node = ggml_graph_node(gf, node_idx);
-            for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
-                if (node->src[src_idx] != nullptr) {
-                    work_stack.push(node->src[src_idx]);
-                }
-            }
-        }
-
-        if (!internal_nodes.empty()) {
-            segment.internal_node_indices.assign(internal_nodes.begin(), internal_nodes.end());
-        }
-
-        std::sort(input_refs.begin(),
-                  input_refs.end(),
-                  [](const Segment::InputRef& a, const Segment::InputRef& b) {
-                      if (a.type != b.type) {
-                          return a.type < b.type;
-                      }
-                      return a.display_name < b.display_name;
-                  });
-        segment.input_refs = input_refs;
-        for (const auto& input : input_refs) {
-            ggml_tensor* current_input = input_tensor(gf, input);
-            size_t tensor_bytes        = current_input == nullptr
-                                             ? 0
-                                             : (input.type == Segment::INPUT_PREVIOUS_CUT
-                                                    ? cache_tensor_bytes(current_input)
-                                                    : ggml_nbytes(current_input));
-            switch (input.type) {
-                case Segment::INPUT_PREVIOUS_CUT:
-                    segment.input_previous_cut_bytes += tensor_bytes;
-                    break;
-                case Segment::INPUT_PARAM:
-                    segment.input_param_bytes += tensor_bytes;
-                    break;
-                case Segment::INPUT_EXTERNAL:
-                default:
-                    segment.input_external_bytes += tensor_bytes;
-                    break;
-            }
-        }
-        for (int output_node_index : segment.output_node_indices) {
-            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
-            segment.output_bytes += cache_tensor_bytes(output);
-        }
-        segment.compute_buffer_size = measure_segment_compute_buffer(backend, gf, segment, log_desc);
-
-        for (int output_node_index : segment.output_node_indices) {
-            available_cut_output_node_indices.insert(output_node_index);
-        }
-        plan.segments.push_back(std::move(segment));
-    }
-
-    bool is_graph_cut_tensor(const ggml_tensor* tensor) {
-        if (tensor == nullptr || tensor->name[0] == '\0') {
-            return false;
-        }
-        return std::strncmp(tensor->name, GGML_RUNNER_CUT_PREFIX, std::strlen(GGML_RUNNER_CUT_PREFIX)) == 0;
-    }
-
-    std::string make_graph_cut_name(const std::string& group, const std::string& output) {
-        return std::string(GGML_RUNNER_CUT_PREFIX) + group + "|" + output;
-    }
-
-    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output) {
-        if (tensor == nullptr) {
-            return;
-        }
-        auto name = make_graph_cut_name(group, output);
-        ggml_set_name(tensor, name.c_str());
-    }
-
-    int leaf_count(ggml_cgraph* gf) {
-        GGML_ASSERT(gf != nullptr);
-        return gf->n_leafs;
-    }
-
-    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index) {
-        GGML_ASSERT(gf != nullptr);
-        if (leaf_index < 0 || leaf_index >= gf->n_leafs) {
-            return nullptr;
-        }
-        return gf->leafs[leaf_index];
-    }
-
-    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor) {
-        if (tensor == nullptr) {
-            return nullptr;
-        }
-        return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
-    }
-
-    ggml_tensor* cache_source_tensor(ggml_tensor* tensor) {
-        if (tensor == nullptr) {
-            return nullptr;
-        }
-        if (tensor_buffer(tensor) == nullptr && tensor->src[0] != nullptr &&
-            ggml_nelements(tensor->src[0]) == ggml_nelements(tensor) &&
-            ggml_nbytes(tensor->src[0]) == ggml_nbytes(tensor)) {
-            return cache_source_tensor(tensor->src[0]);
-        }
-        return tensor->view_src ? tensor->view_src : tensor;
-    }
-
-    size_t cache_tensor_bytes(const ggml_tensor* tensor) {
-        if (tensor == nullptr) {
-            return 0;
-        }
-        const ggml_tensor* cache_src = tensor->view_src ? tensor->view_src : tensor;
-        return ggml_nbytes(cache_src);
-    }
-
-    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) {
-        GGML_ASSERT(gf != nullptr);
-        if (ggml_graph_n_nodes(gf) != plan.n_nodes || gf->n_leafs != plan.n_leafs) {
-            return false;
-        }
-        for (const auto& input_shape_ref : plan.input_shapes) {
-            if (input_shape_ref.leaf_index < 0 || input_shape_ref.leaf_index >= gf->n_leafs) {
-                return false;
-            }
-            ggml_tensor* leaf = gf->leafs[input_shape_ref.leaf_index];
-            if (leaf == nullptr || input_shape_ref.type != leaf->type) {
-                return false;
-            }
-            for (int d = 0; d < GGML_MAX_DIMS; ++d) {
-                if (input_shape_ref.ne[static_cast<size_t>(d)] != leaf->ne[d]) {
-                    return false;
-                }
-            }
-        }
-        return true;
-    }
-
-    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index) {
-        GGML_ASSERT(gf != nullptr);
-        if (output_index >= segment.output_node_indices.size()) {
-            return nullptr;
-        }
-        int node_index = segment.output_node_indices[output_index];
-        if (node_index < 0 || node_index >= ggml_graph_n_nodes(gf)) {
-            return nullptr;
-        }
-        return ggml_graph_node(gf, node_index);
-    }
-
-    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref) {
-        GGML_ASSERT(gf != nullptr);
-        if (input_ref.type == Segment::INPUT_PREVIOUS_CUT) {
-            if (input_ref.node_index < 0 || input_ref.node_index >= ggml_graph_n_nodes(gf)) {
-                return nullptr;
-            }
-            return ggml_graph_node(gf, input_ref.node_index);
-        }
-        if (input_ref.leaf_index < 0 || input_ref.leaf_index >= gf->n_leafs) {
-            return nullptr;
-        }
-        return leaf_tensor(gf, input_ref.leaf_index);
-    }
-
-    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment) {
-        GGML_ASSERT(gf != nullptr);
-        std::vector<ggml_tensor*> tensors;
-        std::unordered_set<ggml_tensor*> seen_tensors;
-        tensors.reserve(segment.input_refs.size());
-        seen_tensors.reserve(segment.input_refs.size());
-        for (const auto& input_ref : segment.input_refs) {
-            if (input_ref.type != Segment::INPUT_PARAM) {
-                continue;
-            }
-            ggml_tensor* tensor = input_tensor(gf, input_ref);
-            if (tensor == nullptr) {
-                continue;
-            }
-            if (seen_tensors.insert(tensor).second) {
-                tensors.push_back(tensor);
-            }
-        }
-        return tensors;
-    }
-
-    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc) {
-        std::vector<ggml_tensor*> tensors = param_tensors(gf, segment);
-        std::vector<ggml_tensor*> filtered_tensors;
-        filtered_tensors.reserve(tensors.size());
-        for (ggml_tensor* tensor : tensors) {
-            if (tensor_buffer(tensor) == nullptr) {
-                LOG_WARN("%s graph cut skipping param input without buffer: segment=%s tensor=%s",
-                         log_desc == nullptr ? "unknown" : log_desc,
-                         segment.group_name.c_str(),
-                         tensor->name);
-                continue;
-            }
-            filtered_tensors.push_back(tensor);
-        }
-        return filtered_tensors;
-    }
-
-    std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
-                                                               const Plan& plan,
-                                                               size_t current_segment_index) {
-        GGML_ASSERT(gf != nullptr);
-        std::unordered_set<std::string> future_input_names;
-        for (size_t seg_idx = current_segment_index + 1; seg_idx < plan.segments.size(); ++seg_idx) {
-            const auto& segment = plan.segments[seg_idx];
-            for (const auto& input_ref : segment.input_refs) {
-                if (input_ref.type != Segment::INPUT_PREVIOUS_CUT) {
-                    continue;
-                }
-                ggml_tensor* current_input = input_tensor(gf, input_ref);
-                if (current_input != nullptr && current_input->name[0] != '\0') {
-                    future_input_names.insert(current_input->name);
-                }
-            }
-        }
-        return future_input_names;
-    }
-
-    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
-                                     const Segment& segment,
-                                     ggml_context** graph_ctx_out) {
-        GGML_ASSERT(gf != nullptr);
-        GGML_ASSERT(graph_ctx_out != nullptr);
-
-        const size_t graph_size = segment.internal_node_indices.size() + segment.input_refs.size() + 8;
-        ggml_init_params params = {
-            /*.mem_size   =*/ggml_graph_overhead_custom(graph_size, false) + 1024,
-            /*.mem_buffer =*/nullptr,
-            /*.no_alloc   =*/true,
-        };
-        ggml_context* graph_ctx = ggml_init(params);
-        GGML_ASSERT(graph_ctx != nullptr);
-        ggml_cgraph* segment_graph = ggml_new_graph_custom(graph_ctx, graph_size, false);
-        GGML_ASSERT(segment_graph != nullptr);
-
-        for (const auto& input : segment.input_refs) {
-            ggml_tensor* current_input = input_tensor(gf, input);
-            if (current_input == nullptr) {
-                continue;
-            }
-            GGML_ASSERT(segment_graph->n_leafs < segment_graph->size);
-            segment_graph->leafs[segment_graph->n_leafs++] = current_input;
-        }
-
-        for (int output_node_index : segment.output_node_indices) {
-            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
-            if (output == nullptr) {
-                continue;
-            }
-            ggml_set_output(output);
-        }
-        for (int node_idx : segment.internal_node_indices) {
-            ggml_graph_add_node(segment_graph, ggml_graph_node(gf, node_idx));
-        }
-        *graph_ctx_out = graph_ctx;
-        return segment_graph;
-    }
-
-    size_t measure_segment_compute_buffer(ggml_backend_t backend,
-                                          ggml_cgraph* gf,
-                                          const Segment& segment,
-                                          const char* log_desc) {
-        GGML_ASSERT(backend != nullptr);
-        GGML_ASSERT(gf != nullptr);
-        if (segment.internal_node_indices.empty()) {
-            return 0;
-        }
-
-        ggml_context* graph_ctx    = nullptr;
-        ggml_cgraph* segment_graph = build_segment_graph(gf, segment, &graph_ctx);
-        ggml_gallocr_t allocr      = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
-
-        size_t sizes[1] = {0};
-        ggml_gallocr_reserve_n_size(
-            allocr,
-            segment_graph,
-            nullptr,
-            nullptr,
-            sizes);
-        size_t buffer_size = sizes[0];
-
-        ggml_gallocr_free(allocr);
-        ggml_free(graph_ctx);
-        return buffer_size;
-    }
-
-    Plan build_plan(ggml_backend_t backend,
-                    ggml_cgraph* gf,
-                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                    const char* log_desc) {
-        GGML_ASSERT(backend != nullptr);
-        GGML_ASSERT(gf != nullptr);
-        Plan plan;
-        plan.available    = true;
-        const int n_nodes = ggml_graph_n_nodes(gf);
-        if (n_nodes <= 0) {
-            return plan;
-        }
-        plan.n_nodes = n_nodes;
-        plan.n_leafs = gf->n_leafs;
-        for (int i = 0; i < gf->n_leafs; ++i) {
-            ggml_tensor* leaf = gf->leafs[i];
-            if (is_params_tensor(params_tensor_set, leaf)) {
-                continue;
-            }
-            auto shape       = input_shape(leaf);
-            shape.leaf_index = i;
-            plan.input_shapes.push_back(shape);
-        }
-
-        std::unordered_map<const ggml_tensor*, int> producer_index;
-        producer_index.reserve(static_cast<size_t>(n_nodes));
-        for (int i = 0; i < n_nodes; ++i) {
-            producer_index[ggml_graph_node(gf, i)] = i;
-        }
-
-        std::vector<Segment> grouped_segments;
-        std::unordered_map<std::string, size_t> group_to_segment;
-        for (int i = 0; i < n_nodes; ++i) {
-            ggml_tensor* node = ggml_graph_node(gf, i);
-            if (!is_graph_cut_tensor(node)) {
-                continue;
-            }
-
-            plan.has_cuts = true;
-            std::string full_name(node->name);
-            std::string payload = full_name.substr(std::strlen(GGML_RUNNER_CUT_PREFIX));
-            size_t sep          = payload.find('|');
-            std::string group   = sep == std::string::npos ? payload : payload.substr(0, sep);
-
-            auto it = group_to_segment.find(group);
-            if (it == group_to_segment.end()) {
-                Segment segment;
-                segment.group_name = group;
-                segment.output_node_indices.push_back(i);
-                group_to_segment[group] = grouped_segments.size();
-                grouped_segments.push_back(std::move(segment));
-            } else {
-                auto& segment = grouped_segments[it->second];
-                segment.output_node_indices.push_back(i);
-            }
-        }
-
-        if (!plan.has_cuts) {
-            return plan;
-        }
-
-        std::unordered_set<int> available_cut_output_node_indices;
-        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
-        for (auto& segment : grouped_segments) {
-            build_segment(gf,
-                          plan,
-                          segment,
-                          producer_index,
-                          available_cut_output_node_indices,
-                          backend,
-                          params_tensor_set,
-                          log_desc);
-        }
-
-        int final_output_index = graph_node_index_by_name(gf, "ggml_runner_final_result_tensor");
-        if (final_output_index < 0) {
-            final_output_index = n_nodes - 1;
-        }
-        ggml_tensor* final_output = final_output_index >= 0 ? ggml_graph_node(gf, final_output_index) : nullptr;
-        if (final_output != nullptr && available_cut_output_node_indices.find(final_output_index) == available_cut_output_node_indices.end()) {
-            Segment final_segment;
-            final_segment.group_name = "ggml_runner.final";
-            final_segment.output_node_indices.push_back(final_output_index);
-            build_segment(gf,
-                          plan,
-                          final_segment,
-                          producer_index,
-                          available_cut_output_node_indices,
-                          backend,
-                          params_tensor_set,
-                          log_desc);
-        }
-
-        return plan;
-    }
-
-    Plan apply_max_vram_budget(ggml_cgraph* gf,
-                               const Plan& base_plan,
-                               size_t max_graph_vram_bytes,
-                               ggml_backend_t backend,
-                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                               const char* log_desc) {
-        GGML_ASSERT(backend != nullptr);
-        GGML_ASSERT(gf != nullptr);
-        int64_t t_budget_begin = ggml_time_ms();
-        if (max_graph_vram_bytes == 0 || !base_plan.has_cuts || base_plan.segments.size() <= 1) {
-            return base_plan;
-        }
-
-        const int n_nodes = ggml_graph_n_nodes(gf);
-        std::unordered_map<const ggml_tensor*, int> producer_index;
-        producer_index.reserve(static_cast<size_t>(n_nodes));
-        for (int i = 0; i < n_nodes; ++i) {
-            producer_index[ggml_graph_node(gf, i)] = i;
-        }
-
-        Plan merged_plan;
-        merged_plan.available = true;
-        merged_plan.has_cuts  = base_plan.has_cuts;
-        merged_plan.valid     = base_plan.valid;
-        merged_plan.n_nodes   = base_plan.n_nodes;
-        merged_plan.n_leafs   = base_plan.n_leafs;
-
-        std::unordered_set<int> available_cut_output_node_indices;
-        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
-
-        size_t start_segment_index = 0;
-        while (start_segment_index < base_plan.segments.size()) {
-            Plan single_plan;
-            auto single_available_cut_output_node_indices = available_cut_output_node_indices;
-            auto single_seed                              = make_segment_seed(base_plan,
-                                                                              start_segment_index,
-                                                                              start_segment_index);
-            build_segment(gf,
-                          single_plan,
-                          single_seed,
-                          producer_index,
-                          single_available_cut_output_node_indices,
-                          backend,
-                          params_tensor_set,
-                          log_desc);
-            GGML_ASSERT(!single_plan.segments.empty());
-
-            size_t best_end_segment_index = start_segment_index;
-            bool can_merge_next_segment   = graph_cut_segment_vram_bytes(single_plan.segments.back()) <= max_graph_vram_bytes;
-
-            while (can_merge_next_segment && best_end_segment_index + 1 < base_plan.segments.size()) {
-                const size_t next_end_segment_index = best_end_segment_index + 1;
-                Plan candidate_plan;
-                auto candidate_available_cut_output_node_indices = available_cut_output_node_indices;
-                auto candidate_seed                              = make_segment_seed(base_plan,
-                                                                                     start_segment_index,
-                                                                                     next_end_segment_index);
-                build_segment(gf,
-                              candidate_plan,
-                              candidate_seed,
-                              producer_index,
-                              candidate_available_cut_output_node_indices,
-                              backend,
-                              params_tensor_set,
-                              log_desc);
-                GGML_ASSERT(!candidate_plan.segments.empty());
-
-                const auto& candidate_segment = candidate_plan.segments.back();
-                if (graph_cut_segment_vram_bytes(candidate_segment) > max_graph_vram_bytes) {
-                    break;
-                }
-
-                best_end_segment_index = next_end_segment_index;
-            }
-
-            auto best_seed = make_segment_seed(base_plan,
-                                               start_segment_index,
-                                               best_end_segment_index);
-            build_segment(gf,
-                          merged_plan,
-                          best_seed,
-                          producer_index,
-                          available_cut_output_node_indices,
-                          backend,
-                          params_tensor_set,
-                          log_desc);
-            start_segment_index = best_end_segment_index + 1;
-        }
-
-        if (log_desc != nullptr && merged_plan.segments.size() != base_plan.segments.size()) {
-            LOG_INFO("%s graph cut max_vram=%.2f MB merged %zu segments -> %zu segments",
-                     log_desc,
-                     max_graph_vram_bytes / 1024.0 / 1024.0,
-                     base_plan.segments.size(),
-                     merged_plan.segments.size());
-        }
-
-        if (log_desc != nullptr) {
-            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
-                     log_desc,
-                     ggml_time_ms() - t_budget_begin);
-        }
-
-        return merged_plan;
-    }
-
-    Plan resolve_plan(ggml_backend_t backend,
-                      ggml_cgraph* gf,
-                      PlanCache* cache,
-                      size_t max_graph_vram_bytes,
-                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                      const char* log_desc) {
-        GGML_ASSERT(backend != nullptr);
-        GGML_ASSERT(gf != nullptr);
-        GGML_ASSERT(cache != nullptr);
-
-        int64_t t_prepare_begin = ggml_time_ms();
-        Plan base_plan;
-        int64_t t_plan_begin = ggml_time_ms();
-        if (cache->graph_cut_plan.available && plan_matches_graph(gf, cache->graph_cut_plan)) {
-            base_plan = cache->graph_cut_plan;
-        } else {
-            base_plan                                = build_plan(backend, gf, params_tensor_set, log_desc);
-            cache->graph_cut_plan                    = base_plan;
-            cache->graph_cut_plan.available          = true;
-            cache->budgeted_graph_cut_plan.available = false;
-            if (log_desc != nullptr) {
-                LOG_INFO("%s build cached graph cut plan done (taking %lld ms)", log_desc, ggml_time_ms() - t_plan_begin);
-            }
-        }
-
-        Plan resolved_plan = base_plan;
-        if (max_graph_vram_bytes > 0 && base_plan.has_cuts) {
-            if (cache->budgeted_graph_cut_plan.available &&
-                cache->budgeted_graph_cut_plan_max_vram_bytes == max_graph_vram_bytes &&
-                plan_matches_graph(gf, cache->budgeted_graph_cut_plan)) {
-                resolved_plan = cache->budgeted_graph_cut_plan;
-            } else {
-                resolved_plan                                 = apply_max_vram_budget(gf,
-                                                                                      base_plan,
-                                                                                      max_graph_vram_bytes,
-                                                                                      backend,
-                                                                                      params_tensor_set,
-                                                                                      log_desc);
-                cache->budgeted_graph_cut_plan                = resolved_plan;
-                cache->budgeted_graph_cut_plan.available      = true;
-                cache->budgeted_graph_cut_plan_max_vram_bytes = max_graph_vram_bytes;
-            }
-        }
-        return resolved_plan;
-    }
-
-}  // namespace sd::ggml_graph_cut

src/ggml_graph_cut.h

@@ -1,104 +0,0 @@
-#ifndef __SD_GGML_GRAPH_CUT_H__
-#define __SD_GGML_GRAPH_CUT_H__
-
-#include <array>
-#include <string>
-#include <unordered_set>
-#include <vector>
-
-#include "ggml-backend.h"
-#include "ggml.h"
-
-namespace sd::ggml_graph_cut {
-
-    struct Segment {
-        enum InputType {
-            INPUT_EXTERNAL = 0,
-            INPUT_PREVIOUS_CUT,
-            INPUT_PARAM,
-        };
-
-        struct InputRef {
-            InputType type = INPUT_EXTERNAL;
-            std::string display_name;
-            int leaf_index = -1;
-            int node_index = -1;
-        };
-
-        size_t compute_buffer_size      = 0;
-        size_t output_bytes             = 0;
-        size_t input_external_bytes     = 0;
-        size_t input_previous_cut_bytes = 0;
-        size_t input_param_bytes        = 0;
-        std::string group_name;
-        std::vector<int> internal_node_indices;
-        std::vector<int> output_node_indices;
-        std::vector<InputRef> input_refs;
-    };
-
-    struct Plan {
-        struct InputShape {
-            int leaf_index                        = -1;
-            ggml_type type                        = GGML_TYPE_COUNT;
-            std::array<int64_t, GGML_MAX_DIMS> ne = {0, 0, 0, 0};
-        };
-
-        bool available = false;
-        bool has_cuts  = false;
-        bool valid     = true;
-        int n_nodes    = 0;
-        int n_leafs    = 0;
-        std::vector<InputShape> input_shapes;
-        std::vector<Segment> segments;
-    };
-
-    struct PlanCache {
-        Plan graph_cut_plan;
-        Plan budgeted_graph_cut_plan;
-        size_t budgeted_graph_cut_plan_max_vram_bytes = 0;
-    };
-
-    static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:";
-
-    bool is_graph_cut_tensor(const ggml_tensor* tensor);
-    std::string make_graph_cut_name(const std::string& group, const std::string& output);
-    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output);
-    int leaf_count(ggml_cgraph* gf);
-    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index);
-    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor);
-    ggml_tensor* cache_source_tensor(ggml_tensor* tensor);
-    size_t cache_tensor_bytes(const ggml_tensor* tensor);
-    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan);
-    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index);
-    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref);
-    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment);
-    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc);
-    std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
-                                                               const Plan& plan,
-                                                               size_t current_segment_index);
-    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
-                                     const Segment& segment,
-                                     ggml_context** graph_ctx_out);
-    size_t measure_segment_compute_buffer(ggml_backend_t backend,
-                                          ggml_cgraph* gf,
-                                          const Segment& segment,
-                                          const char* log_desc);
-    Plan build_plan(ggml_backend_t backend,
-                    ggml_cgraph* gf,
-                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                    const char* log_desc);
-    Plan apply_max_vram_budget(ggml_cgraph* gf,
-                               const Plan& base_plan,
-                               size_t max_graph_vram_bytes,
-                               ggml_backend_t backend,
-                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                               const char* log_desc);
-    Plan resolve_plan(ggml_backend_t backend,
-                      ggml_cgraph* gf,
-                      PlanCache* cache,
-                      size_t max_graph_vram_bytes,
-                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                      const char* log_desc);
-}  // namespace sd::ggml_graph_cut
-
-#endif

src/hidream_o1.hpp

@@ -1,653 +0,0 @@
-#ifndef __SD_HIDREAM_O1_H__
-#define __SD_HIDREAM_O1_H__
-
-#include <algorithm>
-#include <array>
-#include <cmath>
-#include <cstring>
-#include <memory>
-#include <string>
-#include <utility>
-#include <vector>
-
-#include "common_dit.hpp"
-#include "conditioner.hpp"
-#include "llm.hpp"
-#include "util.h"
-
-namespace HiDreamO1 {
-    constexpr int HIDREAM_O1_GRAPH_SIZE = 32768;
-    constexpr int PATCH_SIZE            = 32;
-    constexpr int TIMESTEP_TOKEN_NUM    = 1;
-    constexpr int IMAGE_TOKEN_ID        = 151655;
-    constexpr int VISION_START_TOKEN_ID = 151652;
-
-    static inline std::string repeat_special_token(const std::string& token, int64_t count) {
-        std::string out;
-        out.reserve(static_cast<size_t>(count) * token.size());
-        for (int64_t i = 0; i < count; ++i) {
-            out += token;
-        }
-        return out;
-    }
-
-    static inline std::pair<int, int> calculate_dimensions(int max_size, double ratio) {
-        int width  = static_cast<int>(std::sqrt(max_size * max_size * ratio));
-        int height = static_cast<int>(width / ratio);
-        width      = (width / PATCH_SIZE) * PATCH_SIZE;
-        height     = (height / PATCH_SIZE) * PATCH_SIZE;
-        width      = std::max(width, PATCH_SIZE);
-        height     = std::max(height, PATCH_SIZE);
-        return {width, height};
-    }
-
-    static inline sd::Tensor<float> resize_to_area(const sd::Tensor<float>& image, int image_size) {
-        int64_t width  = image.shape()[0];
-        int64_t height = image.shape()[1];
-        int64_t s_max  = static_cast<int64_t>(image_size) * image_size;
-        double scale   = std::sqrt(static_cast<double>(s_max) / static_cast<double>(width * height));
-
-        std::vector<std::pair<int64_t, int64_t>> sizes = {
-            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
-            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
-            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
-            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
-        };
-        std::sort(sizes.begin(), sizes.end(), [](const auto& a, const auto& b) {
-            return a.first * a.second > b.first * b.second;
-        });
-
-        std::pair<int64_t, int64_t> new_size = sizes.back();
-        for (const auto& size : sizes) {
-            if (size.first > 0 && size.second > 0 && size.first * size.second <= s_max) {
-                new_size = size;
-                break;
-            }
-        }
-
-        double s1 = static_cast<double>(width) / static_cast<double>(new_size.first);
-        double s2 = static_cast<double>(height) / static_cast<double>(new_size.second);
-        sd::Tensor<float> resized;
-        if (s1 < s2) {
-            int64_t resized_h = static_cast<int64_t>(std::llround(height / s1));
-            resized           = sd::ops::interpolate(image,
-                                                     {new_size.first, resized_h, image.shape()[2], image.shape()[3]},
-                                                     sd::ops::InterpolateMode::Bicubic);
-            int64_t top       = (resized_h - new_size.second) / 2;
-            resized           = sd::ops::slice(resized, 1, top, top + new_size.second);
-        } else {
-            int64_t resized_w = static_cast<int64_t>(std::llround(width / s2));
-            resized           = sd::ops::interpolate(image,
-                                                     {resized_w, new_size.second, image.shape()[2], image.shape()[3]},
-                                                     sd::ops::InterpolateMode::Bicubic);
-            int64_t left      = (resized_w - new_size.first) / 2;
-            resized           = sd::ops::slice(resized, 0, left, left + new_size.first);
-        }
-        return resized;
-    }
-
-    static inline std::vector<int32_t> build_position_ids(const std::vector<int32_t>& input_ids,
-                                                          const std::vector<std::array<int32_t, 3>>& image_grids,
-                                                          const std::vector<int32_t>& skip_vision_start_token) {
-        std::vector<int32_t> position_ids(4 * input_ids.size(), 0);
-        int image_index = 0;
-        int st          = 0;
-        int fix_point   = 4096;
-        std::vector<int32_t> out_t;
-        std::vector<int32_t> out_h;
-        std::vector<int32_t> out_w;
-
-        while (st < static_cast<int>(input_ids.size())) {
-            int ed = st;
-            while (ed < static_cast<int>(input_ids.size()) && input_ids[ed] != IMAGE_TOKEN_ID) {
-                ed++;
-            }
-
-            if (ed >= static_cast<int>(input_ids.size())) {
-                int st_idx = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
-                for (int i = 0; i < static_cast<int>(input_ids.size()) - st; ++i) {
-                    out_t.push_back(st_idx + i);
-                    out_h.push_back(st_idx + i);
-                    out_w.push_back(st_idx + i);
-                }
-                break;
-            }
-
-            int text_len = std::max(0, ed - st - skip_vision_start_token[image_index]);
-            int st_idx   = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
-            for (int i = 0; i < text_len; ++i) {
-                out_t.push_back(st_idx + i);
-                out_h.push_back(st_idx + i);
-                out_w.push_back(st_idx + i);
-            }
-
-            auto grid = image_grids[image_index];
-            int base;
-            if (skip_vision_start_token[image_index]) {
-                if (fix_point > 0) {
-                    base      = fix_point;
-                    fix_point = 0;
-                } else {
-                    base = st_idx;
-                }
-            } else {
-                base = text_len + st_idx;
-            }
-            for (int32_t ti = 0; ti < grid[0]; ++ti) {
-                for (int32_t hi = 0; hi < grid[1]; ++hi) {
-                    for (int32_t wi = 0; wi < grid[2]; ++wi) {
-                        out_t.push_back(base + ti);
-                        out_h.push_back(base + hi);
-                        out_w.push_back(base + wi);
-                    }
-                }
-            }
-
-            st = ed + grid[0] * grid[1] * grid[2];
-            image_index++;
-        }
-
-        GGML_ASSERT(out_t.size() == input_ids.size());
-        for (size_t i = 0; i < input_ids.size(); ++i) {
-            // ggml IMROPE consumes 4 flattened position streams:
-            //   [t, h, w, e]
-            // llama.cpp's generic Qwen-VL fallback expands text positions as
-            // [pos, pos, pos, 0]. Keep the extra stream zeroed here too.
-            position_ids[i]                        = out_t[i];
-            position_ids[input_ids.size() + i]     = out_h[i];
-            position_ids[input_ids.size() * 2 + i] = out_w[i];
-            position_ids[input_ids.size() * 3 + i] = 0;
-        }
-        return position_ids;
-    }
-
-    struct TimestepEmbedder : public GGMLBlock {
-        int frequency_embedding_size = 256;
-
-        TimestepEmbedder(int64_t hidden_size) {
-            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true);
-            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
-            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
-            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
-            auto emb   = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10000, 1000.0f);
-            emb        = mlp_0->forward(ctx, emb);
-            emb        = ggml_silu_inplace(ctx->ggml_ctx, emb);
-            emb        = mlp_2->forward(ctx, emb);
-            return emb;
-        }
-    };
-
-    struct BottleneckPatchEmbed : public GGMLBlock {
-        BottleneckPatchEmbed(int64_t in_dim, int64_t pca_dim, int64_t embed_dim) {
-            blocks["proj1"] = std::make_shared<Linear>(in_dim, pca_dim, false);
-            blocks["proj2"] = std::make_shared<Linear>(pca_dim, embed_dim, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            auto proj1 = std::dynamic_pointer_cast<Linear>(blocks["proj1"]);
-            auto proj2 = std::dynamic_pointer_cast<Linear>(blocks["proj2"]);
-            return proj2->forward(ctx, proj1->forward(ctx, x));
-        }
-    };
-
-    struct FinalLayer : public GGMLBlock {
-        FinalLayer(int64_t hidden_size, int64_t out_dim) {
-            blocks["linear"] = std::make_shared<Linear>(hidden_size, out_dim, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
-            return linear->forward(ctx, x);
-        }
-    };
-
-    struct HiDreamO1Params {
-        LLM::LLMParams llm;
-        int patch_size = PATCH_SIZE;
-    };
-
-    static inline HiDreamO1Params make_hidream_o1_params() {
-        HiDreamO1Params params;
-        params.llm.arch                           = LLM::LLMArch::QWEN3_VL;
-        params.llm.hidden_size                    = 4096;
-        params.llm.intermediate_size              = 12288;
-        params.llm.num_layers                     = 36;
-        params.llm.num_heads                      = 32;
-        params.llm.num_kv_heads                   = 8;
-        params.llm.head_dim                       = 128;
-        params.llm.qkv_bias                       = false;
-        params.llm.qk_norm                        = true;
-        params.llm.vocab_size                     = 151936;
-        params.llm.rms_norm_eps                   = 1e-6f;
-        params.llm.vision.arch                    = LLM::LLMVisionArch::QWEN3_VL;
-        params.llm.vision.num_layers              = 27;
-        params.llm.vision.hidden_size             = 1152;
-        params.llm.vision.intermediate_size       = 4304;
-        params.llm.vision.num_heads               = 16;
-        params.llm.vision.out_hidden_size         = 4096;
-        params.llm.vision.patch_size              = 16;
-        params.llm.vision.spatial_merge_size      = 2;
-        params.llm.vision.temporal_patch_size     = 2;
-        params.llm.vision.num_position_embeddings = 2304;
-        return params;
-    }
-
-    struct HiDreamO1Model : public GGMLBlock {
-        HiDreamO1Params params;
-
-        HiDreamO1Model() = default;
-        explicit HiDreamO1Model(HiDreamO1Params params)
-            : params(std::move(params)) {
-            blocks["language_model"] = std::make_shared<LLM::TextModel>(this->params.llm);
-            blocks["t_embedder1"]    = std::make_shared<TimestepEmbedder>(this->params.llm.hidden_size);
-            blocks["x_embedder"]     = std::make_shared<BottleneckPatchEmbed>(this->params.patch_size * this->params.patch_size * 3,
-                                                                          this->params.llm.hidden_size / 4,
-                                                                          this->params.llm.hidden_size);
-            blocks["final_layer2"]   = std::make_shared<FinalLayer>(this->params.llm.hidden_size,
-                                                                  this->params.patch_size * this->params.patch_size * 3);
-        }
-
-        std::shared_ptr<LLM::TextModel> text_model() {
-            return std::dynamic_pointer_cast<LLM::TextModel>(blocks["language_model"]);
-        }
-
-        std::shared_ptr<TimestepEmbedder> timestep_embedder() {
-            return std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder1"]);
-        }
-
-        std::shared_ptr<BottleneckPatchEmbed> patch_embedder() {
-            return std::dynamic_pointer_cast<BottleneckPatchEmbed>(blocks["x_embedder"]);
-        }
-
-        std::shared_ptr<FinalLayer> final_layer() {
-            return std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer2"]);
-        }
-    };
-
-    struct HiDreamO1VisionRunner : public GGMLRunner {
-        HiDreamO1Params params;
-        std::shared_ptr<LLM::VisionModel> model;
-
-        std::vector<int> window_index_vec;
-        std::vector<int> window_inverse_index_vec;
-        std::vector<float> window_mask_vec;
-        std::vector<float> pe_vec;
-        std::array<std::vector<int32_t>, 4> pos_embed_idx_data_;
-        std::array<std::vector<float>, 4> pos_embed_weight_data_;
-
-        HiDreamO1VisionRunner(ggml_backend_t backend,
-                              bool offload_params_to_cpu,
-                              const String2TensorStorage& tensor_storage_map = {},
-                              const std::string& prefix                      = "model.visual")
-            : GGMLRunner(backend, offload_params_to_cpu),
-              params(make_hidream_o1_params()),
-              model(std::make_shared<LLM::VisionModel>(false, params.llm.vision)) {
-            model->init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "hidream_o1_vision";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix = "model.visual") {
-            model->get_param_tensors(tensors, prefix);
-        }
-
-        ggml_tensor* encode_image(GGMLRunnerContext* runner_ctx, ggml_tensor* image) {
-            return LLM::LLMRunner::encode_image_common(this,
-                                                       compute_ctx,
-                                                       runner_ctx,
-                                                       image,
-                                                       params.llm.vision,
-                                                       model,
-                                                       window_index_vec,
-                                                       window_inverse_index_vec,
-                                                       window_mask_vec,
-                                                       pe_vec,
-                                                       pos_embed_idx_data_,
-                                                       pos_embed_weight_data_);
-        }
-
-        ggml_cgraph* build_graph(const sd::Tensor<float>& image_tensor) {
-            ggml_cgraph* gf    = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
-            ggml_tensor* image = make_input(image_tensor);
-            auto runner_ctx    = get_context();
-            auto image_embeds  = encode_image(&runner_ctx, image);
-            ggml_build_forward_expand(gf, image_embeds);
-            return gf;
-        }
-
-        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& image) {
-            auto get_graph = [&]() {
-                return build_graph(image);
-            };
-            auto output = GGMLRunner::compute<float>(get_graph, n_threads, false);
-            return output.has_value() ? std::move(output.value()) : sd::Tensor<float>();
-        }
-    };
-
-    struct HiDreamO1Runner : public GGMLRunner {
-        HiDreamO1Params params;
-        HiDreamO1Model model;
-
-        std::vector<float> attention_mask_vec;
-
-        HiDreamO1Runner(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
-                        const String2TensorStorage& tensor_storage_map = {},
-                        const std::string& prefix                      = "model")
-            : GGMLRunner(backend, offload_params_to_cpu),
-              params(make_hidream_o1_params()) {
-            model = HiDreamO1Model(params);
-            model.init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "hidream_o1";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
-            model.get_param_tensors(tensors, prefix);
-        }
-
-        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
-                                 const sd::Tensor<float>& timestep_tensor,
-                                 const sd::Tensor<int32_t>& input_ids_tensor,
-                                 const sd::Tensor<int32_t>& input_pos_tensor,
-                                 const sd::Tensor<int32_t>& token_types_tensor,
-                                 const sd::Tensor<int32_t>& vinput_mask_tensor,
-                                 const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds_tensor,
-                                 const std::vector<sd::Tensor<float>>& ref_images) {
-            ggml_cgraph* gf        = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
-            ggml_tensor* x         = make_input(x_tensor);
-            ggml_tensor* timestep  = make_input(timestep_tensor);
-            ggml_tensor* input_ids = make_input(input_ids_tensor);
-            ggml_tensor* input_pos = make_input(input_pos_tensor);
-
-            auto text_model   = model.text_model();
-            auto t_embedder1  = model.timestep_embedder();
-            auto x_embedder   = model.patch_embedder();
-            auto final_layer2 = model.final_layer();
-
-            std::vector<ggml_tensor*> ref_image_tensors;
-            for (const auto& image : ref_images) {
-                ref_image_tensors.push_back(make_input(image));
-            }
-
-            attention_mask_vec    = std::vector<float>(static_cast<size_t>(token_types_tensor.shape()[0] * token_types_tensor.shape()[0]), 0.0f);
-            int64_t total_seq_len = token_types_tensor.shape()[0];
-            for (int64_t query = 0; query < total_seq_len; ++query) {
-                bool is_gen = token_types_tensor.values()[static_cast<size_t>(query)] > 0;
-                for (int64_t key = 0; key < total_seq_len; ++key) {
-                    if (!is_gen && key > query) {
-                        attention_mask_vec[static_cast<size_t>(query * total_seq_len + key)] = -INFINITY;
-                    }
-                }
-            }
-            auto attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, total_seq_len, total_seq_len);
-            set_backend_tensor_data(attention_mask, attention_mask_vec.data());
-
-            auto runner_ctx = get_context();
-            auto txt        = text_model->embed(&runner_ctx, input_ids);
-            std::vector<std::pair<int, ggml_tensor*>> image_embeds;
-            image_embeds.reserve(image_embeds_tensor.size());
-            for (const auto& image_embed : image_embeds_tensor) {
-                image_embeds.emplace_back(image_embed.first, make_input(image_embed.second));
-            }
-            txt = LLM::splice_image_embeds(&runner_ctx, txt, image_embeds);
-
-            auto t_emb          = t_embedder1->forward(&runner_ctx, timestep);
-            int64_t txt_seq_len = input_ids->ne[0];
-            if (txt_seq_len > 1) {
-                auto prefix = ggml_ext_slice(compute_ctx, txt, 1, 0, txt_seq_len - 1);
-                txt         = ggml_concat(compute_ctx, prefix, ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1), 1);
-            } else {
-                txt = ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1);
-            }
-
-            auto vinputs          = DiT::pad_and_patchify(&runner_ctx, x, PATCH_SIZE, PATCH_SIZE);
-            int64_t target_tokens = vinputs->ne[1];
-            for (ggml_tensor* ref_image : ref_image_tensors) {
-                auto ref = DiT::pad_and_patchify(&runner_ctx, ref_image, PATCH_SIZE, PATCH_SIZE);
-                vinputs  = ggml_concat(compute_ctx, vinputs, ref, 1);
-            }
-            auto vis = x_embedder->forward(&runner_ctx, vinputs);
-
-            auto inputs_embeds = ggml_concat(compute_ctx, txt, vis, 1);
-            auto hidden_states = text_model->forward_embeds(&runner_ctx, inputs_embeds, input_pos, attention_mask, {});
-            auto x_pred_all    = final_layer2->forward(&runner_ctx, hidden_states);
-
-            int64_t x_pred_start = txt_seq_len;
-            if (!vinput_mask_tensor.empty()) {
-                int64_t seq_len      = static_cast<int64_t>(vinput_mask_tensor.shape()[0]);
-                int64_t first_vinput = 0;
-                while (first_vinput < seq_len && vinput_mask_tensor.values()[static_cast<size_t>(first_vinput)] == 0) {
-                    first_vinput++;
-                }
-                x_pred_start = first_vinput;
-            }
-            auto x_pred = ggml_ext_slice(compute_ctx, x_pred_all, 1, x_pred_start, x_pred_start + target_tokens);
-            x_pred      = DiT::unpatchify_and_crop(compute_ctx, x_pred, x->ne[1], x->ne[0], PATCH_SIZE, PATCH_SIZE);
-
-            float sigma = 1.0f - timestep_tensor.values()[0];
-            sigma       = std::max(1e-6f, sigma);
-            auto out    = ggml_scale(compute_ctx, ggml_sub(compute_ctx, x, x_pred), 1.0f / sigma);
-
-            ggml_build_forward_expand(gf, out);
-            return gf;
-        }
-
-        sd::Tensor<float> compute(int n_threads,
-                                  const sd::Tensor<float>& x,
-                                  const sd::Tensor<float>& timestep,
-                                  const sd::Tensor<int32_t>& input_ids,
-                                  const sd::Tensor<int32_t>& input_pos,
-                                  const sd::Tensor<int32_t>& token_types,
-                                  const sd::Tensor<int32_t>& vinput_mask,
-                                  const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
-                                  const std::vector<sd::Tensor<float>>& ref_images) {
-            auto get_graph = [&]() {
-                return build_graph(x, timestep, input_ids, input_pos, token_types, vinput_mask, image_embeds, ref_images);
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
-        }
-    };
-
-    struct HiDreamO1Conditioner : public Conditioner {
-        Qwen2Tokenizer tokenizer;
-        std::shared_ptr<HiDreamO1VisionRunner> vision_runner;
-
-        HiDreamO1Conditioner(ggml_backend_t backend,
-                             bool offload_params_to_cpu,
-                             const String2TensorStorage& tensor_storage_map = {})
-            : vision_runner(std::make_shared<HiDreamO1VisionRunner>(backend, offload_params_to_cpu, tensor_storage_map)) {}
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-            vision_runner->get_param_tensors(tensors);
-        }
-
-        void alloc_params_buffer() override {
-            vision_runner->alloc_params_buffer();
-        }
-
-        void free_params_buffer() override {
-            vision_runner->free_params_buffer();
-        }
-
-        size_t get_params_buffer_size() override {
-            return vision_runner->get_params_buffer_size();
-        }
-
-        void set_max_graph_vram_bytes(size_t max_graph_vram_bytes) override {
-            vision_runner->set_max_graph_vram_bytes(max_graph_vram_bytes);
-        }
-
-        void set_flash_attention_enabled(bool enabled) override {
-            vision_runner->set_flash_attention_enabled(enabled);
-        }
-
-        void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-            vision_runner->set_weight_adapter(adapter);
-        }
-
-        SDCondition get_learned_condition(int n_threads,
-                                          const ConditionerParams& conditioner_params) override {
-            SDCondition result;
-
-            int width                = conditioner_params.width;
-            int height               = conditioner_params.height;
-            int64_t target_image_len = static_cast<int64_t>(width / PATCH_SIZE) * static_cast<int64_t>(height / PATCH_SIZE);
-
-            std::vector<sd::Tensor<float>> ref_images;
-            if (conditioner_params.ref_images != nullptr) {
-                ref_images = *conditioner_params.ref_images;
-            }
-
-            std::vector<std::pair<int, sd::Tensor<float>>> vlm_images;
-            std::vector<std::array<int32_t, 3>> image_grids;
-            std::vector<int32_t> skip_vision_start;
-
-            std::string prompt = "<|im_start|>user\n";
-
-            if (ref_images.empty()) {
-                prompt += conditioner_params.text;
-                prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
-                auto input_ids = tokenizer.encode(prompt, nullptr);
-
-                std::vector<int32_t> input_ids_pad = input_ids;
-                input_ids_pad.push_back(VISION_START_TOKEN_ID);
-                input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
-
-                image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
-                skip_vision_start.push_back(1);
-
-                std::vector<int32_t> token_types(input_ids_pad.size(), 0);
-                int txt_seq_len = static_cast<int>(input_ids.size());
-                int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
-                for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
-                    token_types[i] = 1;
-                }
-
-                auto position_ids = build_position_ids(input_ids_pad, image_grids, skip_vision_start);
-
-                std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
-                std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
-                std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
-                std::vector<int32_t> vinput_mask(token_types.size(), 0);
-                for (int64_t i = txt_seq_len; i < static_cast<int64_t>(vinput_mask.size()); ++i) {
-                    vinput_mask[static_cast<size_t>(i)] = 1;
-                }
-                std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
-
-                result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
-                result.c_position_ids = sd::Tensor<int32_t>(position_shape, position_ids);
-                result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
-                result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
-                return result;
-            }
-
-            int K = static_cast<int>(ref_images.size());
-            int max_size;
-            if (K == 1) {
-                max_size = std::max(height, width);
-            } else if (K == 2) {
-                max_size = std::max(height, width) * 48 / 64;
-            } else if (K <= 4) {
-                max_size = std::max(height, width) / 2;
-            } else if (K <= 8) {
-                max_size = std::max(height, width) * 24 / 64;
-            } else {
-                max_size = std::max(height, width) / 4;
-            }
-
-            int cond_img_size;
-            if (K <= 4) {
-                cond_img_size = 384;
-            } else if (K <= 8) {
-                cond_img_size = 384 * 48 / 64;
-            } else {
-                cond_img_size = 384 / 2;
-            }
-
-            for (const auto& ref_image : ref_images) {
-                auto resized_ref = resize_to_area(ref_image, max_size);
-                resized_ref      = sd::ops::clamp(resized_ref, 0.0f, 1.0f);
-
-                // VLM image: Qwen3-VL expects mean=[0.5]/std=[0.5] (i.e. range [-1,1]),
-                // not CLIP normalization. Resize the already-resized ref directly to
-                // (cond_w, cond_h) to match the Python pipeline's pil_r.resize().
-                auto dims                   = calculate_dimensions(cond_img_size,
-                                                                   static_cast<double>(resized_ref.shape()[0]) / static_cast<double>(resized_ref.shape()[1]));
-                sd::Tensor<float> vlm_image = sd::ops::interpolate(
-                    resized_ref,
-                    {dims.first, dims.second, resized_ref.shape()[2], resized_ref.shape()[3]});
-                vlm_image            = vlm_image * 2.0f - 1.0f;
-                int64_t image_tokens = static_cast<int64_t>(dims.first / PATCH_SIZE) * static_cast<int64_t>(dims.second / PATCH_SIZE);
-
-                auto patch_img = resized_ref * 2.0f - 1.0f;
-                result.c_ref_images.push_back(std::move(patch_img));
-                int64_t prompt_start = static_cast<int64_t>(tokenizer.encode(prompt + "<|vision_start|>", nullptr).size());
-                prompt += "<|vision_start|>";
-                prompt += repeat_special_token("<|image_pad|>", image_tokens);
-                prompt += "<|vision_end|>";
-                vlm_images.emplace_back(static_cast<int>(prompt_start), std::move(vlm_image));
-                image_grids.push_back({1, dims.second / PATCH_SIZE, dims.first / PATCH_SIZE});
-                skip_vision_start.push_back(0);
-            }
-
-            prompt += conditioner_params.text;
-            prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
-            auto input_ids = tokenizer.encode(prompt, nullptr);
-
-            std::vector<int32_t> input_ids_pad = input_ids;
-            input_ids_pad.push_back(VISION_START_TOKEN_ID);
-            input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
-            image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
-            skip_vision_start.push_back(1);
-
-            for (const auto& ref_image : result.c_ref_images) {
-                int64_t ref_len = static_cast<int64_t>(ref_image.shape()[0] / PATCH_SIZE) * static_cast<int64_t>(ref_image.shape()[1] / PATCH_SIZE);
-                input_ids_pad.push_back(VISION_START_TOKEN_ID);
-                input_ids_pad.insert(input_ids_pad.end(), ref_len - 1, IMAGE_TOKEN_ID);
-                image_grids.push_back({1, static_cast<int32_t>(ref_image.shape()[1] / PATCH_SIZE), static_cast<int32_t>(ref_image.shape()[0] / PATCH_SIZE)});
-                skip_vision_start.push_back(1);
-            }
-
-            std::vector<int32_t> token_types(input_ids_pad.size(), 0);
-            int txt_seq_len = static_cast<int>(input_ids.size());
-            int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
-            for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
-                token_types[i] = 1;
-            }
-
-            std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
-            std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
-            std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
-            std::vector<int32_t> vinput_mask(token_types.size(), 0);
-            for (int i = txt_seq_len; i < static_cast<int>(vinput_mask.size()); ++i) {
-                vinput_mask[static_cast<size_t>(i)] = 1;
-            }
-            std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
-
-            result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
-            result.c_position_ids = sd::Tensor<int32_t>(position_shape, build_position_ids(input_ids_pad, image_grids, skip_vision_start));
-            result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
-            result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
-            result.c_image_embeds.reserve(vlm_images.size());
-            for (const auto& vlm_image : vlm_images) {
-                auto image_embed = vision_runner->compute(n_threads, vlm_image.second);
-                if (image_embed.empty()) {
-                    LOG_ERROR("hidream_o1 conditioner: encode VLM image failed");
-                    return SDCondition();
-                }
-                result.c_image_embeds.emplace_back(vlm_image.first, std::move(image_embed));
-            }
-            return result;
-        }
-    };
-}  // namespace HiDreamO1
-
-#endif  // __SD_HIDREAM_O1_H__

src/lora.hpp

@@ -129,7 +129,7 @@ struct LoraModel : public GGMLRunner {
         }
     }
 
-    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
+    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -152,17 +152,17 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lora_up_name);
             if (iter != lora_tensors.end()) {
-                lora_up = ggml_ext_cast_f32(ctx, backend, iter->second);
+                lora_up = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(lora_mid_name);
             if (iter != lora_tensors.end()) {
-                lora_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
+                lora_mid = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(lora_down_name);
             if (iter != lora_tensors.end()) {
-                lora_down = ggml_ext_cast_f32(ctx, backend, iter->second);
+                lora_down = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             if (lora_up == nullptr || lora_down == nullptr) {
@@ -208,7 +208,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
+    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -225,7 +225,7 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(diff_name);
             if (iter != lora_tensors.end()) {
-                curr_updown = ggml_ext_cast_f32(ctx, backend, iter->second);
+                curr_updown = ggml_ext_cast_f32(ctx, iter->second);
             } else {
                 break;
             }
@@ -248,7 +248,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
+    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -276,33 +276,33 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(hada_1_down_name);
             if (iter != lora_tensors.end()) {
-                hada_1_down = ggml_ext_cast_f32(ctx, backend, iter->second);
+                hada_1_down = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_up_name);
             if (iter != lora_tensors.end()) {
-                hada_1_up = ggml_ext_cast_f32(ctx, backend, iter->second);
+                hada_1_up = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_1_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
+                hada_1_mid = ggml_ext_cast_f32(ctx, iter->second);
                 hada_1_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_1_up));
             }
 
             iter = lora_tensors.find(hada_2_down_name);
             if (iter != lora_tensors.end()) {
-                hada_2_down = ggml_ext_cast_f32(ctx, backend, iter->second);
+                hada_2_down = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_up_name);
             if (iter != lora_tensors.end()) {
-                hada_2_up = ggml_ext_cast_f32(ctx, backend, iter->second);
+                hada_2_up = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_2_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
+                hada_2_mid = ggml_ext_cast_f32(ctx, iter->second);
                 hada_2_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_2_up));
             }
 
@@ -351,7 +351,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
+    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -378,24 +378,24 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lokr_w1_name);
             if (iter != lora_tensors.end()) {
-                lokr_w1 = ggml_ext_cast_f32(ctx, backend, iter->second);
+                lokr_w1 = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             iter = lora_tensors.find(lokr_w2_name);
             if (iter != lora_tensors.end()) {
-                lokr_w2 = ggml_ext_cast_f32(ctx, backend, iter->second);
+                lokr_w2 = ggml_ext_cast_f32(ctx, iter->second);
             }
 
             int64_t rank = 1;
             if (lokr_w1 == nullptr) {
                 iter = lora_tensors.find(lokr_w1_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_a = ggml_ext_cast_f32(ctx, backend, iter->second);
+                    lokr_w1_a = ggml_ext_cast_f32(ctx, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w1_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_b = ggml_ext_cast_f32(ctx, backend, iter->second);
+                    lokr_w1_b = ggml_ext_cast_f32(ctx, iter->second);
                 }
 
                 if (lokr_w1_a == nullptr || lokr_w1_b == nullptr) {
@@ -410,12 +410,12 @@ struct LoraModel : public GGMLRunner {
             if (lokr_w2 == nullptr) {
                 iter = lora_tensors.find(lokr_w2_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_a = ggml_ext_cast_f32(ctx, backend, iter->second);
+                    lokr_w2_a = ggml_ext_cast_f32(ctx, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w2_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_b = ggml_ext_cast_f32(ctx, backend, iter->second);
+                    lokr_w2_b = ggml_ext_cast_f32(ctx, iter->second);
                 }
 
                 if (lokr_w2_a == nullptr || lokr_w2_b == nullptr) {
@@ -468,23 +468,23 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_backend_t backend, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
+    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
         // lora
         ggml_tensor* diff = nullptr;
         if (with_lora_and_lokr) {
-            diff = get_lora_weight_diff(model_tensor_name, ctx, backend);
+            diff = get_lora_weight_diff(model_tensor_name, ctx);
         }
         // diff
         if (diff == nullptr) {
-            diff = get_raw_weight_diff(model_tensor_name, ctx, backend);
+            diff = get_raw_weight_diff(model_tensor_name, ctx);
         }
         // loha
         if (diff == nullptr) {
-            diff = get_loha_weight_diff(model_tensor_name, ctx, backend);
+            diff = get_loha_weight_diff(model_tensor_name, ctx);
         }
         // lokr
         if (diff == nullptr && with_lora_and_lokr) {
-            diff = get_lokr_weight_diff(model_tensor_name, ctx, backend);
+            diff = get_lokr_weight_diff(model_tensor_name, ctx);
         }
         if (diff != nullptr) {
             if (ggml_nelements(diff) < ggml_nelements(model_tensor)) {
@@ -502,7 +502,6 @@ struct LoraModel : public GGMLRunner {
     }
 
     ggml_tensor* get_out_diff(ggml_context* ctx,
-                              ggml_backend_t backend,
                               ggml_tensor* x,
                               WeightAdapter::ForwardParams forward_params,
                               const std::string& model_tensor_name) {
@@ -591,7 +590,7 @@ struct LoraModel : public GGMLRunner {
                 }
                 scale_value *= multiplier;
 
-                auto curr_out_diff = ggml_ext_lokr_forward(ctx, backend, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
+                auto curr_out_diff = ggml_ext_lokr_forward(ctx, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
                 if (out_diff == nullptr) {
                     out_diff = curr_out_diff;
                 } else {
@@ -762,7 +761,7 @@ struct LoraModel : public GGMLRunner {
             ggml_tensor* model_tensor     = it.second;
 
             // lora
-            ggml_tensor* diff = get_weight_diff(model_tensor_name, runtime_backend, compute_ctx, model_tensor);
+            ggml_tensor* diff = get_weight_diff(model_tensor_name, compute_ctx, model_tensor);
             if (diff == nullptr) {
                 continue;
             }
@@ -775,7 +774,7 @@ struct LoraModel : public GGMLRunner {
 
             ggml_tensor* final_tensor;
             if (model_tensor->type != GGML_TYPE_F32 && model_tensor->type != GGML_TYPE_F16) {
-                final_tensor = ggml_ext_cast_f32(compute_ctx, runtime_backend, model_tensor);
+                final_tensor = ggml_ext_cast_f32(compute_ctx, model_tensor);
                 final_tensor = ggml_add_inplace(compute_ctx, final_tensor, diff);
                 final_tensor = ggml_cpy(compute_ctx, final_tensor, model_tensor);
             } else {
@@ -842,35 +841,34 @@ public:
         : lora_models(lora_models) {
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
         for (auto& lora_model : lora_models) {
-            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, backend, ctx, weight, with_lora_and_lokr);
+            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora_and_lokr);
             if (diff == nullptr) {
                 continue;
             }
 
             if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
-                weight = ggml_ext_cast_f32(ctx, backend, weight);
+                weight = ggml_ext_cast_f32(ctx, weight);
             }
             weight = ggml_add(ctx, weight, diff);
         }
         return weight;
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name) override {
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name) override {
-        return patch_weight(ctx, backend, weight, weight_name, true);
+        return patch_weight(ctx, weight, weight_name, true);
     }
 
     ggml_tensor* forward_with_lora(ggml_context* ctx,
-                                   ggml_backend_t backend,
                                    ggml_tensor* x,
                                    ggml_tensor* w,
                                    ggml_tensor* b,
                                    const std::string& prefix,
                                    WeightAdapter::ForwardParams forward_params) override {
-        w = patch_weight(ctx, backend, w, prefix + "weight", false);
+        w = patch_weight(ctx, w, prefix + "weight", false);
         if (b) {
-            b = patch_weight(ctx, backend, b, prefix + "bias", false);
+            b = patch_weight(ctx, b, prefix + "bias", false);
         }
         ggml_tensor* out;
         if (forward_params.op_type == ForwardParams::op_type_t::OP_LINEAR) {
@@ -892,7 +890,7 @@ public:
                                    forward_params.conv2d.scale);
         }
         for (auto& lora_model : lora_models) {
-            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, backend, x, forward_params, prefix + "weight");
+            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, x, forward_params, prefix + "weight");
             if (out_diff == nullptr) {
                 continue;
             }

src/mmdit.hpp

@@ -767,8 +767,6 @@ public:
             auto context_x = block->forward(ctx, context, x, c_mod);
             context        = context_x.first;
             x              = context_x.second;
-            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.joint_blocks." + std::to_string(i), "context");
-            sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.joint_blocks." + std::to_string(i), "x");
         }
 
         x = final_layer->forward(ctx, x, c_mod);  // (N, T, patch_size ** 2 * out_channels)
@@ -811,11 +809,6 @@ public:
 
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
         }
-        sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.prelude", "x");
-        sd::ggml_graph_cut::mark_graph_cut(c, "mmdit.prelude", "c");
-        if (context != nullptr) {
-            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.prelude", "context");
-        }
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 

src/model.cpp

@@ -23,11 +23,24 @@
 
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
+#include "ggml-cpu.h"
 #include "ggml.h"
-#include "ggml_extend_backend.hpp"
 #include "zip.h"
 
 #include "name_conversion.h"
+#include "stable-diffusion.h"
+
+#ifdef SD_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef SD_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#ifdef SD_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
 
 /*================================================= Preprocess ==================================================*/
 
@@ -68,7 +81,7 @@ const char* unused_tensors[] = {
     "first_stage_model.bn.",
 };
 
-bool is_unused_tensor(const std::string& name) {
+bool is_unused_tensor(std::string name) {
     for (size_t i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
         if (starts_with(name, unused_tensors[i])) {
             return true;
@@ -437,10 +450,6 @@ SDVersion ModelLoader::get_sd_version() {
         if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
             return VERSION_SD3;
         }
-        if (tensor_storage.name.find("model.x_embedder.proj1.weight") != std::string::npos &&
-            tensor_storage_map.find("model.language_model.layers.0.self_attn.q_proj.weight") != tensor_storage_map.end()) {
-            return VERSION_HIDREAM_O1;
-        }
         if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
             return VERSION_QWEN_IMAGE;
         }
@@ -678,8 +687,8 @@ std::map<ggml_type, uint32_t> ModelLoader::get_vae_wtype_stat() {
     return wtype_stat;
 }
 
-TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules) {
+static std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std::string& tensor_type_rules) {
-    TensorTypeRules result;
+    std::vector<std::pair<std::string, ggml_type>> result;
     for (const auto& item : split_string(tensor_type_rules, ',')) {
         if (item.size() == 0)
             continue;
@@ -734,168 +743,8 @@ void ModelLoader::set_wtype_override(ggml_type wtype, std::string tensor_type_ru
     }
 }
 
-void ModelLoader::process_model_files(bool enable_mmap, bool writable_mmap) {
-    if (model_files_processed) {
-        return;
-    }
-
-    int64_t start_time = ggml_time_ms();
-
-    std::vector<TensorStorage> processed_tensor_storages;
-    for (const auto& [name, tensor_storage] : tensor_storage_map) {
-        if (is_unused_tensor(tensor_storage.name)) {
-            continue;
-        }
-        processed_tensor_storages.push_back(tensor_storage);
-    }
-
-    for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
-        std::string file_path = file_paths_[file_index];
-
-        std::vector<TensorStorage> file_tensors;
-        for (const auto& ts : processed_tensor_storages) {
-            if (ts.file_index == file_index) {
-                file_tensors.push_back(ts);
-            }
-        }
-        if (file_tensors.empty()) {
-            continue;
-        }
-
-        bool is_zip = false;
-        for (auto const& ts : file_tensors) {
-            if (ts.index_in_zip >= 0) {
-                is_zip = true;
-                break;
-            }
-        }
-
-        ModelFileData fdata = {};
-        fdata.path          = file_path;
-        fdata.is_zip        = is_zip;
-        fdata.tensors       = std::move(file_tensors);
-
-        if (enable_mmap && !is_zip) {
-            LOG_DEBUG("using mmap for I/O");
-            std::unique_ptr<MmapWrapper> mmapped = MmapWrapper::create(file_path, writable_mmap);
-            if (mmapped) {
-                uint8_t* mmap_data             = static_cast<uint8_t*>(mmapped->writable_data());
-                ggml_backend_buffer_t buf_mmap = ggml_backend_cpu_buffer_from_ptr(mmap_data, mmapped->size());
-                if (buf_mmap) {
-                    LOG_INFO("using mmap for '%s'", file_path.c_str());
-                    fdata.mmbuffer = std::shared_ptr<struct ggml_backend_buffer>(buf_mmap, ggml_backend_buffer_free);
-                } else {
-                    LOG_WARN("mmap: failed to create backend buffer for file %s", fdata.path.c_str());
-                }
-                fdata.mmapped = std::shared_ptr<MmapWrapper>(std::move(mmapped));
-            } else {
-                LOG_WARN("failed to memory-map '%s' (falling back to read())", file_path.c_str());
-            }
-        } else if (!is_zip) {
-            LOG_INFO("NOT using mmap for '%s' (mmap disabled by caller)",
-                     file_path.c_str());
-        }
-
-        file_data.push_back(std::move(fdata));
-    }
-
-    model_files_processed = true;
-
-    int64_t end_time        = ggml_time_ms();
-    int64_t process_time_ms = end_time - start_time;
-
-    LOG_INFO("model files processing completed in %.2fs", process_time_ms / 1000.f);
-}
-
-std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
-                                                       std::set<std::string> ignore_tensors,
-                                                       bool writable_mmap) {
-    process_model_files(true, writable_mmap);
-
-    std::vector<MmapTensorStore> result;
-    uint64_t mapped_bytes = 0;
-    size_t mapped_tensors = 0;
-
-    LOG_DEBUG("memory-mapping tensors...");
-
-    int64_t t_start = ggml_time_ms();
-
-    for (auto& fdata : file_data) {
-        if (!fdata.mmbuffer)
-            continue;
-
-        const std::vector<TensorStorage>& file_tensors = fdata.tensors;
-
-        size_t file_mapped_bytes   = 0;
-        size_t file_mapped_tensors = 0;
-
-        for (const auto& tensor_storage : file_tensors) {
-            const std::string& name = tensor_storage.name;
-
-            bool is_ignored = false;
-            for (const auto& ignore_prefix : ignore_tensors) {
-                if (starts_with(name, ignore_prefix)) {
-                    is_ignored = true;
-                    break;
-                }
-            }
-            if (is_ignored)
-                continue;
-
-            auto it = tensors.find(name);
-            if (it == tensors.end())
-                continue;
-
-            ggml_tensor* dst_tensor = it->second;
-            if (dst_tensor == nullptr)
-                continue;
-
-            if (tensor_storage.type != dst_tensor->type)
-                continue;
-
-            size_t tensor_size   = tensor_storage.nbytes();
-            size_t tensor_offset = tensor_storage.offset;
-
-            if (tensor_storage.ne[0] != dst_tensor->ne[0] ||
-                tensor_storage.ne[1] != dst_tensor->ne[1] ||
-                tensor_storage.ne[2] != dst_tensor->ne[2] ||
-                tensor_storage.ne[3] != dst_tensor->ne[3] ||
-                tensor_size != ggml_nbytes(dst_tensor)) {
-                // let load_tensors worry about this
-                continue;
-            }
-
-            ggml_backend_buffer_t buf_mmap = fdata.mmbuffer.get();
-            uint8_t* mmap_data             = static_cast<uint8_t*>(ggml_backend_buffer_get_base(buf_mmap));
-            dst_tensor->buffer             = buf_mmap;
-            dst_tensor->data               = mmap_data + tensor_offset;
-
-            file_mapped_bytes += tensor_size;
-            file_mapped_tensors++;
-        }
-
-        if (file_mapped_bytes > 0) {
-            mapped_tensors += file_mapped_tensors;
-            mapped_bytes += file_mapped_bytes;
-            result.push_back({fdata.mmapped, fdata.mmbuffer});
-        }
-    }
-
-    int64_t t_end       = ggml_time_ms();
-    int64_t duration_ms = t_end - t_start;
-
-    LOG_INFO("memory-mapped %zu tensors in %zu files (%.2f MB), taking %.2fs",
-             mapped_tensors,
-             result.size(),
-             mapped_bytes / (1024.0 * 1024.0),
-             duration_ms / 1000.0);
-
-    return result;
-}
-
 bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
-    process_model_files(enable_mmap, false);
+    int64_t process_time_ms = 0;
-
     std::atomic<int64_t> read_time_ms(0);
     std::atomic<int64_t> memcpy_time_ms(0);
     std::atomic<int64_t> copy_to_backend_time_ms(0);
@@ -907,25 +756,52 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
 
     int64_t start_time = ggml_time_ms();
 
-    size_t total_tensors_to_process = 0;
+    std::vector<TensorStorage> processed_tensor_storages;
-    for (const auto& fdata : file_data) {
+    for (const auto& [name, tensor_storage] : tensor_storage_map) {
-        total_tensors_to_process += fdata.tensors.size();
+        if (is_unused_tensor(tensor_storage.name)) {
+            continue;
         }
+        processed_tensor_storages.push_back(tensor_storage);
+    }
+
+    process_time_ms = ggml_time_ms() - start_time;
 
     bool success                          = true;
     size_t total_tensors_processed        = 0;
-    const int64_t t_start          = start_time;
+    const size_t total_tensors_to_process = processed_tensor_storages.size();
+    const int64_t t_start                 = ggml_time_ms();
     int last_n_threads                    = 1;
 
-    for (auto& fdata : file_data) {
+    for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
-        const std::string& file_path = fdata.path;
+        std::string file_path = file_paths_[file_index];
         LOG_DEBUG("loading tensors from %s", file_path.c_str());
 
-        const std::vector<TensorStorage>& file_tensors = fdata.tensors;
+        std::vector<const TensorStorage*> file_tensors;
+        for (const auto& ts : processed_tensor_storages) {
+            if (ts.file_index == file_index) {
+                file_tensors.push_back(&ts);
+            }
+        }
+        if (file_tensors.empty()) {
+            continue;
+        }
 
-        bool is_zip = fdata.is_zip;
+        bool is_zip = false;
+        for (auto const& ts : file_tensors) {
+            if (ts->index_in_zip >= 0) {
+                is_zip = true;
+                break;
+            }
+        }
 
-        std::shared_ptr<MmapWrapper> mmapped = fdata.mmapped;
+        std::unique_ptr<MmapWrapper> mmapped;
+        if (enable_mmap && !is_zip) {
+            LOG_DEBUG("using mmap for I/O");
+            mmapped = MmapWrapper::create(file_path);
+            if (!mmapped) {
+                LOG_WARN("failed to memory-map '%s'", file_path.c_str());
+            }
+        }
 
         int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
         if (n_threads < 1) {
@@ -967,7 +843,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         break;
                     }
 
-                    const TensorStorage& tensor_storage = file_tensors[idx];
+                    const TensorStorage& tensor_storage = *file_tensors[idx];
                     ggml_tensor* dst_tensor             = nullptr;
 
                     t0 = ggml_time_ms();
@@ -984,11 +860,6 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         continue;
                     }
 
-                    // skip mmapped tensors
-                    if (dst_tensor->buffer != nullptr && dst_tensor->buffer == fdata.mmbuffer.get()) {
-                        continue;
-                    }
-
                     size_t nbytes_to_read = tensor_storage.nbytes_to_read();
 
                     auto read_data = [&](char* buf, size_t n) {
@@ -1132,8 +1003,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     }
 
     int64_t end_time = ggml_time_ms();
-    LOG_INFO("loading tensors completed, taking %.2fs (read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
+    LOG_INFO("loading tensors completed, taking %.2fs (process: %.2fs, read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
              (end_time - start_time) / 1000.f,
+             process_time_ms / 1000.f,
              (read_time_ms.load() / (float)last_n_threads) / 1000.f,
              (memcpy_time_ms.load() / (float)last_n_threads) / 1000.f,
              (convert_time_ms.load() / (float)last_n_threads) / 1000.f,
@@ -1249,6 +1121,91 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
     return false;
 }
 
+bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules_str) {
+    auto tensor_type_rules = parse_tensor_type_rules(tensor_type_rules_str);
+    auto get_tensor_type   = [&](const TensorStorage& tensor_storage) -> ggml_type {
+        const std::string& name = tensor_storage.name;
+        ggml_type tensor_type   = tensor_storage.type;
+        ggml_type dst_type      = type;
+
+        for (const auto& tensor_type_rule : tensor_type_rules) {
+            std::regex pattern(tensor_type_rule.first);
+            if (std::regex_search(name, pattern)) {
+                dst_type = tensor_type_rule.second;
+                break;
+            }
+        }
+
+        if (tensor_should_be_converted(tensor_storage, dst_type)) {
+            tensor_type = dst_type;
+        }
+
+        return tensor_type;
+    };
+
+    auto backend    = ggml_backend_cpu_init();
+    size_t mem_size = 1 * 1024 * 1024;  // for padding
+    mem_size += tensor_storage_map.size() * ggml_tensor_overhead();
+    mem_size += get_params_mem_size(backend, type);
+    LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
+    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
+
+    if (ggml_ctx == nullptr) {
+        LOG_ERROR("ggml_init failed for GGUF writer");
+        ggml_backend_free(backend);
+        return false;
+    }
+
+    std::vector<ggml_tensor*> tensors;
+    std::mutex tensor_mutex;
+    auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+        const std::string& name = tensor_storage.name;
+        ggml_type tensor_type   = get_tensor_type(tensor_storage);
+
+        std::lock_guard<std::mutex> lock(tensor_mutex);
+        ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
+        if (tensor == nullptr) {
+            LOG_ERROR("ggml_new_tensor failed");
+            return false;
+        }
+        ggml_set_name(tensor, name.c_str());
+
+        // LOG_DEBUG("%s %d %s %d[%d %d %d %d] %d[%d %d %d %d]", name.c_str(),
+        // ggml_nbytes(tensor), ggml_type_name(tensor_type),
+        // tensor_storage.n_dims,
+        // tensor_storage.ne[0], tensor_storage.ne[1], tensor_storage.ne[2], tensor_storage.ne[3],
+        // tensor->n_dims, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+
+        if (!tensor->data) {
+            GGML_ASSERT(ggml_nelements(tensor) == 0);
+            // avoid crashing the gguf writer by setting a dummy pointer for zero-sized tensors
+            LOG_DEBUG("setting dummy pointer for zero-sized tensor %s", name.c_str());
+            tensor->data = ggml_get_mem_buffer(ggml_ctx);
+        }
+
+        *dst_tensor = tensor;
+        tensors.push_back(tensor);
+
+        return true;
+    };
+
+    bool success = load_tensors(on_new_tensor_cb);
+    ggml_backend_free(backend);
+    LOG_INFO("load tensors done");
+
+    std::string error;
+    if (success) {
+        success = write_gguf_file(file_path, tensors, &error);
+    }
+
+    if (!success && !error.empty()) {
+        LOG_ERROR("%s", error.c_str());
+    }
+
+    ggml_free(ggml_ctx);
+    return success;
+}
+
 int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type) {
     size_t alignment = 128;
     if (backend != nullptr) {
@@ -1268,3 +1225,28 @@ int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type)
 
     return mem_size;
 }
+
+bool convert(const char* input_path,
+             const char* vae_path,
+             const char* output_path,
+             sd_type_t output_type,
+             const char* tensor_type_rules,
+             bool convert_name) {
+    ModelLoader model_loader;
+
+    if (!model_loader.init_from_file(input_path)) {
+        LOG_ERROR("init model loader from file failed: '%s'", input_path);
+        return false;
+    }
+
+    if (vae_path != nullptr && strlen(vae_path) > 0) {
+        if (!model_loader.init_from_file(vae_path, "vae.")) {
+            LOG_ERROR("init model loader from file failed: '%s'", vae_path);
+            return false;
+        }
+    }
+    if (convert_name) {
+        model_loader.convert_tensors_name();
+    }
+    return model_loader.save_to_gguf_file(output_path, (ggml_type)output_type, tensor_type_rules);
+}

src/model.h

@@ -42,7 +42,6 @@ enum SDVersion {
     VERSION_ANIMA,
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
-    VERSION_HIDREAM_O1,
     VERSION_Z_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
@@ -164,7 +163,6 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_sd3(version) ||
         sd_version_is_wan(version) ||
         sd_version_is_qwen_image(version) ||
-        version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
         sd_version_is_ernie_image(version)) {
@@ -191,31 +189,11 @@ enum PMVersion {
 };
 
 typedef OrderedMap<std::string, TensorStorage> String2TensorStorage;
-using TensorTypeRules = std::vector<std::pair<std::string, ggml_type>>;
-
-TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);
-
-class MmapWrapper;
-
-struct ModelFileData {
-    std::string path;
-    std::vector<TensorStorage> tensors;
-    std::shared_ptr<MmapWrapper> mmapped;
-    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
-    bool is_zip;
-};
-
-struct MmapTensorStore {
-    std::shared_ptr<MmapWrapper> mmapped;
-    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
-};
 
 class ModelLoader {
 protected:
     SDVersion version_ = VERSION_COUNT;
     std::vector<std::string> file_paths_;
-    std::vector<ModelFileData> file_data;
-    bool model_files_processed = false;
     String2TensorStorage tensor_storage_map;
 
     void add_tensor_storage(const TensorStorage& tensor_storage);
@@ -239,10 +217,6 @@ public:
     std::map<ggml_type, uint32_t> get_vae_wtype_stat();
     String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
     void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
-    void process_model_files(bool enable_mmap = false, bool writable_mmap = true);
-    std::vector<MmapTensorStore> mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
-                                              std::set<std::string> ignore_tensors = {},
-                                              bool writable                        = true);
     bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0, bool use_mmap = false);
     bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                       std::set<std::string> ignore_tensors = {},
@@ -257,6 +231,7 @@ public:
         return names;
     }
 
+    bool save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules);
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;

src/model_io/gguf_io.cpp

@@ -95,7 +95,7 @@ bool read_gguf_file(const std::string& file_path,
 }
 
 bool write_gguf_file(const std::string& file_path,
-                     const std::vector<TensorWriteInfo>& tensors,
+                     const std::vector<ggml_tensor*>& tensors,
                      std::string* error) {
     gguf_context* gguf_ctx = gguf_init_empty();
     if (gguf_ctx == nullptr) {
@@ -103,8 +103,7 @@ bool write_gguf_file(const std::string& file_path,
         return false;
     }
 
-    for (const TensorWriteInfo& write_tensor : tensors) {
+    for (ggml_tensor* tensor : tensors) {
-        ggml_tensor* tensor = write_tensor.tensor;
         if (tensor == nullptr) {
             set_error(error, "null tensor cannot be written to GGUF");
             gguf_free(gguf_ctx);

src/model_io/gguf_io.h

@@ -11,7 +11,7 @@ bool read_gguf_file(const std::string& file_path,
                     std::vector<TensorStorage>& tensor_storages,
                     std::string* error = nullptr);
 bool write_gguf_file(const std::string& file_path,
-                     const std::vector<TensorWriteInfo>& tensors,
+                     const std::vector<ggml_tensor*>& tensors,
                      std::string* error = nullptr);
 
 #endif  // __SD_MODEL_IO_GGUF_IO_H__

src/model_io/safetensors_io.cpp

@@ -8,7 +8,6 @@
 
 #include "binary_io.h"
 #include "json.hpp"
-#include "util.h"
 
 static constexpr size_t ST_HEADER_SIZE_LEN = 8;
 
@@ -61,7 +60,7 @@ bool is_safetensors_file(const std::string& file_path) {
     return true;
 }
 
-static ggml_type safetensors_dtype_to_ggml_type(const std::string& dtype) {
+static ggml_type str_to_ggml_type(const std::string& dtype) {
     ggml_type ttype = GGML_TYPE_COUNT;
     if (dtype == "F16") {
         ttype = GGML_TYPE_F16;
@@ -155,7 +154,7 @@ bool read_safetensors_file(const std::string& file_path,
         size_t begin = tensor_info["data_offsets"][0].get<size_t>();
         size_t end   = tensor_info["data_offsets"][1].get<size_t>();
 
-        ggml_type type = safetensors_dtype_to_ggml_type(dtype);
+        ggml_type type = str_to_ggml_type(dtype);
         if (type == GGML_TYPE_COUNT) {
             set_error(error, "unsupported dtype '" + dtype + "' (tensor '" + name + "')");
             return false;
@@ -222,95 +221,3 @@ bool read_safetensors_file(const std::string& file_path,
 
     return true;
 }
-
-static bool ggml_type_to_safetensors_dtype(ggml_type type, std::string* dtype) {
-    switch (type) {
-        case GGML_TYPE_F16:
-            *dtype = "F16";
-            return true;
-        case GGML_TYPE_BF16:
-            *dtype = "BF16";
-            return true;
-        case GGML_TYPE_F32:
-            *dtype = "F32";
-            return true;
-        case GGML_TYPE_I32:
-            *dtype = "I32";
-            return true;
-        default:
-            return false;
-    }
-}
-
-bool write_safetensors_file(const std::string& file_path,
-                            const std::vector<TensorWriteInfo>& tensors,
-                            std::string* error) {
-    nlohmann::ordered_json header = nlohmann::ordered_json::object();
-
-    uint64_t data_offset = 0;
-    for (const TensorWriteInfo& write_tensor : tensors) {
-        ggml_tensor* tensor = write_tensor.tensor;
-        if (tensor == nullptr) {
-            set_error(error, "null tensor cannot be written to safetensors");
-            return false;
-        }
-
-        const std::string name = ggml_get_name(tensor);
-        std::string dtype;
-        if (!ggml_type_to_safetensors_dtype(tensor->type, &dtype)) {
-            set_error(error,
-                      "unsupported safetensors dtype '" + std::string(ggml_type_name(tensor->type)) +
-                          "' for tensor '" + name + "'");
-            return false;
-        }
-
-        const uint64_t tensor_nbytes = ggml_nbytes(tensor);
-
-        nlohmann::ordered_json json_tensor_info = nlohmann::ordered_json::object();
-        json_tensor_info["dtype"]               = dtype;
-
-        nlohmann::ordered_json shape = nlohmann::ordered_json::array();
-        for (int i = 0; i < write_tensor.n_dims; ++i) {
-            shape.push_back(write_tensor.ne[write_tensor.n_dims - 1 - i]);
-        }
-        json_tensor_info["shape"] = shape;
-
-        nlohmann::ordered_json data_offsets = nlohmann::ordered_json::array();
-        data_offsets.push_back(data_offset);
-        data_offsets.push_back(data_offset + tensor_nbytes);
-        json_tensor_info["data_offsets"] = data_offsets;
-
-        header[name] = json_tensor_info;
-        data_offset += tensor_nbytes;
-    }
-
-    const std::string header_str = header.dump();
-
-    std::ofstream file(file_path, std::ios::binary);
-    if (!file.is_open()) {
-        set_error(error, "failed to open '" + file_path + "' for writing");
-        return false;
-    }
-
-    LOG_INFO("trying to save tensors to %s", file_path.c_str());
-    model_io::write_u64(file, header_str.size());
-    file.write(header_str.data(), header_str.size());
-    if (!file) {
-        set_error(error, "failed to write safetensors header to '" + file_path + "'");
-        return false;
-    }
-
-    for (const TensorWriteInfo& write_tensor : tensors) {
-        ggml_tensor* tensor        = write_tensor.tensor;
-        const std::string name     = ggml_get_name(tensor);
-        const size_t tensor_nbytes = ggml_nbytes(tensor);
-        file.write((const char*)tensor->data, tensor_nbytes);
-        if (!file) {
-            set_error(error,
-                      "failed to write tensor '" + name + "' to '" + file_path + "'");
-            return false;
-        }
-    }
-
-    return true;
-}

src/model_io/safetensors_io.h

@@ -10,8 +10,5 @@ bool is_safetensors_file(const std::string& file_path);
 bool read_safetensors_file(const std::string& file_path,
                            std::vector<TensorStorage>& tensor_storages,
                            std::string* error = nullptr);
-bool write_safetensors_file(const std::string& file_path,
-                            const std::vector<TensorWriteInfo>& tensors,
-                            std::string* error = nullptr);
 
 #endif  // __SD_MODEL_IO_SAFETENSORS_IO_H__

src/model_io/tensor_storage.h

@@ -121,12 +121,6 @@ struct TensorStorage {
     }
 };
 
-struct TensorWriteInfo {
-    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
-    int n_dims              = 0;
-    ggml_tensor* tensor     = nullptr;
-};
-
 typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
 
 #endif  // __SD_TENSOR_STORAGE_H__

src/preprocessing.hpp

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

src/qwen_image.hpp

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

src/t5.hpp

@@ -251,8 +251,7 @@ public:
                          ggml_tensor* x,
                          ggml_tensor* past_bias                = nullptr,
                          ggml_tensor* attention_mask           = nullptr,
-                         ggml_tensor* relative_position_bucket = nullptr,
+                         ggml_tensor* relative_position_bucket = nullptr) {
-                         const std::string& graph_cut_prefix   = "") {
         // x: [N, n_token, model_dim]
         for (int i = 0; i < num_layers; i++) {
             auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
@@ -260,9 +259,6 @@ public:
             auto ret  = block->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
             x         = ret.first;
             past_bias = ret.second;
-            if (!graph_cut_prefix.empty()) {
-                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".block." + std::to_string(i), "x");
-            }
         }
 
         auto final_layer_norm = std::dynamic_pointer_cast<T5LayerNorm>(blocks["final_layer_norm"]);
@@ -309,8 +305,7 @@ public:
         auto encoder = std::dynamic_pointer_cast<T5Stack>(blocks["encoder"]);
 
         auto x = shared->forward(ctx, input_ids);
-        sd::ggml_graph_cut::mark_graph_cut(x, "t5.prelude", "x");
+        x      = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
-        x = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket, "t5");
         return x;
     }
 };

src/tensor.hpp

@@ -815,202 +815,11 @@ namespace sd {
     namespace ops {
         enum class InterpolateMode {
             Nearest,
-            NearestExact,
             NearestMax,
             NearestMin,
             NearestAvg,
-            Bilinear,
-            Bicubic,
-            Lanczos,
         };
 
-        inline bool is_nearest_like_interpolate_mode(InterpolateMode mode) {
-            return mode == InterpolateMode::Nearest ||
-                   mode == InterpolateMode::NearestExact ||
-                   mode == InterpolateMode::NearestMax ||
-                   mode == InterpolateMode::NearestMin ||
-                   mode == InterpolateMode::NearestAvg;
-        }
-
-        inline bool is_2d_filter_interpolate_mode(InterpolateMode mode) {
-            return mode == InterpolateMode::Bilinear ||
-                   mode == InterpolateMode::Bicubic ||
-                   mode == InterpolateMode::Lanczos;
-        }
-
-        inline int64_t nearest_exact_interpolate_index(int64_t output_index,
-                                                       int64_t input_size,
-                                                       int64_t output_size) {
-            const double scale  = static_cast<double>(input_size) / static_cast<double>(output_size);
-            const double center = (static_cast<double>(output_index) + 0.5) * scale - 0.5;
-            return std::min(std::max<int64_t>(static_cast<int64_t>(std::floor(center + 0.5)), 0), input_size - 1);
-        }
-
-        inline double linear_interpolate_weight(double x) {
-            x = std::abs(x);
-            return x < 1.0 ? 1.0 - x : 0.0;
-        }
-
-        inline double cubic_interpolate_weight(double x) {
-            constexpr double a = -0.75;  // Match PyTorch bicubic interpolation.
-            x                  = std::abs(x);
-            if (x <= 1.0) {
-                return ((a + 2.0) * x - (a + 3.0)) * x * x + 1.0;
-            }
-            if (x < 2.0) {
-                return ((a * x - 5.0 * a) * x + 8.0 * a) * x - 4.0 * a;
-            }
-            return 0.0;
-        }
-
-        inline double sinc(double x) {
-            constexpr double pi = 3.14159265358979323846;
-            if (std::abs(x) < 1e-12) {
-                return 1.0;
-            }
-            const double pix = pi * x;
-            return std::sin(pix) / pix;
-        }
-
-        inline double lanczos_interpolate_weight(double x) {
-            constexpr double radius = 3.0;
-            x                       = std::abs(x);
-            if (x >= radius) {
-                return 0.0;
-            }
-            return sinc(x) * sinc(x / radius);
-        }
-
-        struct InterpolateContributor {
-            int64_t index;
-            double weight;
-        };
-
-        inline std::vector<std::vector<InterpolateContributor>> make_interpolate_contributors(
-            int64_t input_size,
-            int64_t output_size,
-            InterpolateMode mode,
-            bool antialias) {
-            std::vector<std::vector<InterpolateContributor>> contributors(static_cast<size_t>(output_size));
-            const double scale        = static_cast<double>(input_size) / static_cast<double>(output_size);
-            const double filter_scale = antialias ? std::max(1.0, scale) : 1.0;
-
-            for (int64_t out = 0; out < output_size; ++out) {
-                const double center = (static_cast<double>(out) + 0.5) * scale - 0.5;
-                int64_t start       = 0;
-                int64_t end         = 0;
-
-                if (mode == InterpolateMode::Bilinear) {
-                    const double support = filter_scale;
-                    start                = static_cast<int64_t>(std::ceil(center - support));
-                    end                  = static_cast<int64_t>(std::floor(center + support));
-                } else if (mode == InterpolateMode::Bicubic) {
-                    const double support = 2.0 * filter_scale;
-                    start                = static_cast<int64_t>(std::ceil(center - support));
-                    end                  = static_cast<int64_t>(std::floor(center + support));
-                } else if (mode == InterpolateMode::Lanczos) {
-                    const double support = 3.0 * filter_scale;
-                    start                = static_cast<int64_t>(std::ceil(center - support));
-                    end                  = static_cast<int64_t>(std::floor(center + support));
-                } else {
-                    tensor_throw_invalid_argument("Unsupported 2D filter interpolate mode: mode=" +
-                                                  std::to_string(static_cast<int>(mode)));
-                }
-
-                double weight_sum                                      = 0.0;
-                std::vector<InterpolateContributor>& axis_contributors = contributors[static_cast<size_t>(out)];
-                axis_contributors.reserve(static_cast<size_t>(end - start + 1));
-
-                for (int64_t in = start; in <= end; ++in) {
-                    double weight = 0.0;
-                    if (mode == InterpolateMode::Bilinear) {
-                        weight = linear_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
-                    } else if (mode == InterpolateMode::Bicubic) {
-                        weight = cubic_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
-                    } else {
-                        weight = lanczos_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
-                    }
-
-                    if (weight == 0.0) {
-                        continue;
-                    }
-
-                    const int64_t clamped_index = std::min(std::max<int64_t>(in, 0), input_size - 1);
-                    axis_contributors.push_back({clamped_index, weight});
-                    weight_sum += weight;
-                }
-
-                if ((antialias || mode == InterpolateMode::Lanczos) &&
-                    std::abs(weight_sum) > 1e-12) {
-                    for (auto& contributor : axis_contributors) {
-                        contributor.weight /= weight_sum;
-                    }
-                }
-
-                if (axis_contributors.empty()) {
-                    const int64_t nearest = std::min(
-                        std::max<int64_t>(static_cast<int64_t>(std::floor(center + 0.5)), 0),
-                        input_size - 1);
-                    axis_contributors.push_back({nearest, 1.0});
-                }
-            }
-
-            return contributors;
-        }
-
-        template <typename T>
-        inline Tensor<T> interpolate_2d_filter(const Tensor<T>& input,
-                                               const std::vector<int64_t>& output_shape,
-                                               InterpolateMode mode,
-                                               bool antialias) {
-            if (input.dim() < 2) {
-                tensor_throw_invalid_argument("2D filter interpolate requires rank >= 2: input_shape=" +
-                                              tensor_shape_to_string(input.shape()) + ", output_shape=" +
-                                              tensor_shape_to_string(output_shape));
-            }
-            for (size_t i = 2; i < output_shape.size(); ++i) {
-                if (input.shape()[i] != output_shape[i]) {
-                    tensor_throw_invalid_argument("2D filter interpolate only supports resizing dimensions 0 and 1: input_shape=" +
-                                                  tensor_shape_to_string(input.shape()) + ", output_shape=" +
-                                                  tensor_shape_to_string(output_shape));
-                }
-            }
-
-            Tensor<T> output(output_shape);
-            const int64_t input_width   = input.shape()[0];
-            const int64_t input_height  = input.shape()[1];
-            const int64_t output_width  = output_shape[0];
-            const int64_t output_height = output_shape[1];
-            const int64_t input_plane   = input_width * input_height;
-            const int64_t output_plane  = output_width * output_height;
-            const int64_t plane_count   = input.numel() / input_plane;
-
-            auto x_contributors = make_interpolate_contributors(input_width, output_width, mode, antialias);
-            auto y_contributors = make_interpolate_contributors(input_height, output_height, mode, antialias);
-
-            for (int64_t plane = 0; plane < plane_count; ++plane) {
-                const int64_t input_plane_offset  = plane * input_plane;
-                const int64_t output_plane_offset = plane * output_plane;
-                for (int64_t y = 0; y < output_height; ++y) {
-                    const auto& y_axis = y_contributors[static_cast<size_t>(y)];
-                    for (int64_t x = 0; x < output_width; ++x) {
-                        const auto& x_axis = x_contributors[static_cast<size_t>(x)];
-                        double value       = 0.0;
-                        for (const auto& yc : y_axis) {
-                            const int64_t input_row_offset = input_plane_offset + yc.index * input_width;
-                            for (const auto& xc : x_axis) {
-                                value += static_cast<double>(input.data()[input_row_offset + xc.index]) *
-                                         xc.weight * yc.weight;
-                            }
-                        }
-                        output.data()[output_plane_offset + y * output_width + x] = static_cast<T>(value);
-                    }
-                }
-            }
-
-            return output;
-        }
-
         inline int64_t normalize_slice_bound(int64_t index, int64_t dim_size) {
             if (index < 0) {
                 index += dim_size;
@@ -1205,20 +1014,17 @@ namespace sd {
         inline Tensor<T> interpolate(const Tensor<T>& input,
                                      std::vector<int64_t> output_shape,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false,
+                                     bool align_corners   = false) {
-                                     bool antialias       = false) {
+            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
-            const bool is_nearest_like_mode = is_nearest_like_interpolate_mode(mode);
+                                               mode == InterpolateMode::NearestMax ||
-            const bool is_2d_filter_mode    = is_2d_filter_interpolate_mode(mode);
+                                               mode == InterpolateMode::NearestMin ||
-            if (!is_nearest_like_mode && !is_2d_filter_mode) {
+                                               mode == InterpolateMode::NearestAvg);
-                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+            if (!is_nearest_like_mode) {
-                                              std::to_string(static_cast<int>(mode)));
+                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
-            }
-            if (antialias && !is_2d_filter_mode) {
-                tensor_throw_invalid_argument("Tensor interpolate antialias requires a 2D filter mode: mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for tensor interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()) + ", output_shape=" +
                                               tensor_shape_to_string(output_shape));
             }
@@ -1245,10 +1051,6 @@ namespace sd {
                 }
             }
 
-            if (is_2d_filter_mode) {
-                return interpolate_2d_filter(input, output_shape, mode, antialias);
-            }
-
             bool has_downsampling = false;
             for (int64_t i = 0; i < input.dim(); ++i) {
                 if (input.shape()[i] > output_shape[i]) {
@@ -1258,21 +1060,13 @@ namespace sd {
             }
 
             Tensor<T> output(std::move(output_shape));
-            if (mode == InterpolateMode::Nearest ||
+            if (mode == InterpolateMode::Nearest || !has_downsampling) {
-                mode == InterpolateMode::NearestExact ||
-                !has_downsampling) {
                 for (int64_t flat = 0; flat < output.numel(); ++flat) {
                     std::vector<int64_t> output_coord = tensor_unravel_index(flat, output.shape());
                     std::vector<int64_t> input_coord(static_cast<size_t>(input.dim()), 0);
                     for (size_t i = 0; i < static_cast<size_t>(input.dim()); ++i) {
-                        if (mode == InterpolateMode::NearestExact) {
-                            input_coord[i] = nearest_exact_interpolate_index(output_coord[i],
-                                                                             input.shape()[i],
-                                                                             output.shape()[i]);
-                        } else {
                         input_coord[i] = output_coord[i] * input.shape()[i] / output.shape()[i];
                     }
-                    }
                     output[flat] = input.index(input_coord);
                 }
 
@@ -1289,12 +1083,6 @@ namespace sd {
                         return T(0);
                     case InterpolateMode::Nearest:
                         return T(0);
-                    case InterpolateMode::NearestExact:
-                        return T(0);
-                    case InterpolateMode::Bilinear:
-                    case InterpolateMode::Bicubic:
-                    case InterpolateMode::Lanczos:
-                        break;
                 }
 
                 tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
@@ -1314,12 +1102,6 @@ namespace sd {
                         break;
                     case InterpolateMode::Nearest:
                         break;
-                    case InterpolateMode::NearestExact:
-                        break;
-                    case InterpolateMode::Bilinear:
-                    case InterpolateMode::Bicubic:
-                    case InterpolateMode::Lanczos:
-                        break;
                 }
             };
 
@@ -1375,20 +1157,17 @@ namespace sd {
                                      const std::optional<std::vector<int64_t>>& size,
                                      const std::optional<std::vector<double>>& scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false,
+                                     bool align_corners   = false) {
-                                     bool antialias       = false) {
+            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
-            const bool is_nearest_like_mode = is_nearest_like_interpolate_mode(mode);
+                                               mode == InterpolateMode::NearestMax ||
-            const bool is_2d_filter_mode    = is_2d_filter_interpolate_mode(mode);
+                                               mode == InterpolateMode::NearestMin ||
-            if (!is_nearest_like_mode && !is_2d_filter_mode) {
+                                               mode == InterpolateMode::NearestAvg);
-                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+            if (!is_nearest_like_mode) {
-                                              std::to_string(static_cast<int>(mode)));
+                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
-            }
-            if (antialias && !is_2d_filter_mode) {
-                tensor_throw_invalid_argument("Tensor interpolate antialias requires a 2D filter mode: mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for tensor interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()));
             }
             if (size.has_value() == scale_factor.has_value()) {
@@ -1432,7 +1211,7 @@ namespace sd {
                 }
             }
 
-            return interpolate(input, std::move(output_shape), mode, align_corners, antialias);
+            return interpolate(input, std::move(output_shape), mode, align_corners);
         }
 
         template <typename T>
@@ -1440,14 +1219,12 @@ namespace sd {
                                      const std::optional<std::vector<int64_t>>& size,
                                      double scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false,
+                                     bool align_corners   = false) {
-                                     bool antialias       = false) {
             return interpolate(input,
                                size,
                                std::vector<double>(size.has_value() ? size->size() : input.dim(), scale_factor),
                                mode,
-                               align_corners,
+                               align_corners);
-                               antialias);
         }
 
         template <typename T>

src/tokenizers/clip_tokenizer.cpp

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

src/tokenizers/mistral_tokenizer.cpp

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

src/tokenizers/qwen2_tokenizer.cpp

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

src/unet.hpp

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

src/upscaler.cpp

@@ -1,31 +1,50 @@
-#include "upscaler.h"
+#include "esrgan.hpp"
 #include "ggml_extend.hpp"
 #include "model.h"
 #include "stable-diffusion.h"
 #include "util.h"
 
-UpscalerGGML::UpscalerGGML(int n_threads,
+struct UpscalerGGML {
-                           bool direct,
+    ggml_backend_t backend    = nullptr;  // general backend
-                           int tile_size)
+    ggml_type model_data_type = GGML_TYPE_F16;
+    std::shared_ptr<ESRGAN> esrgan_upscaler;
+    std::string esrgan_path;
+    int n_threads;
+    bool direct   = false;
+    int tile_size = 128;
+
+    UpscalerGGML(int n_threads,
+                 bool direct   = false,
+                 int tile_size = 128)
         : n_threads(n_threads),
           direct(direct),
           tile_size(tile_size) {
     }
 
-void UpscalerGGML::set_max_graph_vram_bytes(size_t max_vram_bytes) {
+    bool load_from_file(const std::string& esrgan_path,
-    max_graph_vram_bytes = max_vram_bytes;
-    if (esrgan_upscaler) {
-        esrgan_upscaler->set_max_graph_vram_bytes(max_vram_bytes);
-    }
-}
-
-bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
                         bool offload_params_to_cpu,
                         int n_threads) {
         ggml_log_set(ggml_log_callback_default, nullptr);
-
+#ifdef SD_USE_CUDA
-    backend = sd_get_default_backend();
+        LOG_DEBUG("Using CUDA backend");
-
+        backend = ggml_backend_cuda_init(0);
+#endif
+#ifdef SD_USE_METAL
+        LOG_DEBUG("Using Metal backend");
+        backend = ggml_backend_metal_init();
+#endif
+#ifdef SD_USE_VULKAN
+        LOG_DEBUG("Using Vulkan backend");
+        backend = ggml_backend_vk_init(0);
+#endif
+#ifdef SD_USE_OPENCL
+        LOG_DEBUG("Using OpenCL backend");
+        backend = ggml_backend_opencl_init();
+#endif
+#ifdef SD_USE_SYCL
+        LOG_DEBUG("Using SYCL backend");
+        backend = ggml_backend_sycl_init(0);
+#endif
         ModelLoader model_loader;
         if (!model_loader.init_from_file_and_convert_name(esrgan_path)) {
             LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
@@ -37,7 +56,6 @@ bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
         }
         LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
         esrgan_upscaler = std::make_shared<ESRGAN>(backend, offload_params_to_cpu, tile_size, model_loader.get_tensor_storage_map());
-    esrgan_upscaler->set_max_graph_vram_bytes(max_graph_vram_bytes);
         if (direct) {
             esrgan_upscaler->set_conv2d_direct_enabled(true);
         }
@@ -47,7 +65,7 @@ bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
         return true;
     }
 
-sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_tensor) {
+    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor) {
         sd::Tensor<float> upscaled;
         if (tile_size <= 0 || (input_tensor.shape()[0] <= tile_size && input_tensor.shape()[1] <= tile_size)) {
             upscaled = esrgan_upscaler->compute(n_threads, input_tensor);
@@ -80,7 +98,7 @@ sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_te
         return upscaled;
     }
 
-sd_image_t UpscalerGGML::upscale(sd_image_t input_image, uint32_t upscale_factor) {
+    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
         // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
         sd_image_t upscaled_image = {0, 0, 0, nullptr};
         int output_width          = (int)input_image.width * esrgan_upscaler->scale;
@@ -101,6 +119,7 @@ sd_image_t UpscalerGGML::upscale(sd_image_t input_image, uint32_t upscale_factor
         upscaled_image = upscaled_data;
         return upscaled_image;
     }
+};
 
 struct upscaler_ctx_t {
     UpscalerGGML* upscaler = nullptr;

src/upscaler.h

@@ -1,33 +0,0 @@
-#ifndef __SD_UPSCALER_H__
-#define __SD_UPSCALER_H__
-
-#include "esrgan.hpp"
-#include "stable-diffusion.h"
-#include "tensor.hpp"
-
-#include <memory>
-#include <string>
-
-struct UpscalerGGML {
-    ggml_backend_t backend    = nullptr;  // general backend
-    ggml_type model_data_type = GGML_TYPE_F16;
-    std::shared_ptr<ESRGAN> esrgan_upscaler;
-    std::string esrgan_path;
-    int n_threads;
-    bool direct                 = false;
-    int tile_size               = 128;
-    size_t max_graph_vram_bytes = 0;
-
-    UpscalerGGML(int n_threads,
-                 bool direct   = false,
-                 int tile_size = 128);
-
-    bool load_from_file(const std::string& esrgan_path,
-                        bool offload_params_to_cpu,
-                        int n_threads);
-    void set_max_graph_vram_bytes(size_t max_vram_bytes);
-    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor);
-    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor);
-};
-
-#endif  // __SD_UPSCALER_H__

src/util.cpp

@@ -23,9 +23,8 @@
 #include <unistd.h>
 #endif
 
-#include "ggml-backend.h"
+#include "ggml-cpu.h"
 #include "ggml.h"
-#include "ggml_extend_backend.hpp"
 #include "stable-diffusion.h"
 
 bool ends_with(const std::string& str, const std::string& ending) {
@@ -112,7 +111,7 @@ private:
     HANDLE hmapping_;
 };
 
-std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bool writable) {
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
     void* mapped_data = nullptr;
     size_t file_size  = 0;
 
@@ -120,10 +119,10 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bo
         filename.c_str(),
         GENERIC_READ,
         FILE_SHARE_READ,
-        nullptr,
+        NULL,
         OPEN_EXISTING,
         FILE_ATTRIBUTE_NORMAL,
-        nullptr);
+        NULL);
 
     if (file_handle == INVALID_HANDLE_VALUE) {
         return nullptr;
@@ -137,20 +136,16 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bo
 
     file_size = static_cast<size_t>(size.QuadPart);
 
-    DWORD page_prot = writable ? PAGE_WRITECOPY : PAGE_READONLY;
+    HANDLE mapping_handle = CreateFileMapping(file_handle, NULL, PAGE_READONLY, 0, 0, NULL);
 
-    HANDLE mapping_handle = CreateFileMapping(file_handle, nullptr, page_prot, 0, 0, nullptr);
+    if (mapping_handle == NULL) {
-
-    if (mapping_handle == nullptr) {
         CloseHandle(file_handle);
         return nullptr;
     }
 
-    DWORD view_access = writable ? FILE_MAP_COPY : FILE_MAP_READ;
+    mapped_data = MapViewOfFile(mapping_handle, FILE_MAP_READ, 0, 0, file_size);
 
-    mapped_data = MapViewOfFile(mapping_handle, view_access, 0, 0, file_size);
+    if (mapped_data == NULL) {
-
-    if (mapped_data == nullptr) {
         CloseHandle(mapping_handle);
         CloseHandle(file_handle);
         return nullptr;
@@ -176,85 +171,28 @@ bool is_directory(const std::string& path) {
     return (stat(path.c_str(), &buffer) == 0 && S_ISDIR(buffer.st_mode));
 }
 
-struct MmapFlags {
-    bool sequential;
-    bool populate;
-    bool willneed;
-    bool dontneed;
-};
-
-static MmapFlags get_mmap_flags() {
-    MmapFlags result          = {};
-    const char* SD_MMAP_FLAGS = std::getenv("SD_MMAP_FLAGS");
-    if (SD_MMAP_FLAGS && *SD_MMAP_FLAGS) {
-        std::stringstream ss(SD_MMAP_FLAGS);
-        std::string token;
-        while (std::getline(ss, token, ',')) {
-            std::string ntoken = trim(token);
-            std::transform(ntoken.begin(), ntoken.end(), ntoken.begin(), ::tolower);
-            if (ntoken == "sequential") {
-                result.sequential = true;
-            } else if (ntoken == "populate") {
-                result.populate = true;
-            } else if (ntoken == "willneed") {
-                result.willneed = true;
-            } else if (ntoken == "dontneed") {
-                result.dontneed = true;
-            }
-        }
-    }
-    return result;
-}
-
 class MmapWrapperImpl : public MmapWrapper {
 public:
-    MmapWrapperImpl(void* data, size_t size, int fd)
+    MmapWrapperImpl(void* data, size_t size)
-        : MmapWrapper(data, size), fd_(fd) {}
+        : MmapWrapper(data, size) {}
 
     ~MmapWrapperImpl() override {
-#ifdef __linux__
-        auto cfg_flags = get_mmap_flags();
-
-        // Drop the kernel pagecache pages for this file. madvise(DONTNEED)
-        // alone only unmaps from the process address space; pagecache
-        // entries persist (`free` reports them as buff/cache and the OOM
-        // killer doesn't touch them, but they ARE counted against
-        // overcommit and can starve other allocations on tight-RAM
-        // systems). posix_fadvise(POSIX_FADV_DONTNEED) is the documented
-        // way to evict pagecache for a specific fd's pages.
-        if (cfg_flags.dontneed) {
-            madvise(data_, size_, MADV_DONTNEED);
-            posix_fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED);
-        }
-#endif
         munmap(data_, size_);
-        close(fd_);
     }
-
-private:
-    int fd_;
 };
 
-std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bool writable) {
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
     int file_descriptor = open(filename.c_str(), O_RDONLY);
     if (file_descriptor == -1) {
         return nullptr;
     }
 
-    auto cfg_flags = get_mmap_flags();
-
     int mmap_flags = MAP_PRIVATE;
 
 #ifdef __linux__
-    // Sequential access hint helps the kernel read-ahead efficiently and
+    // performance flags used by llama.cpp
-    // also encourages eviction of already-read pages (the kernel keeps
+    // posix_fadvise(file_descriptor, 0, 0, POSIX_FADV_SEQUENTIAL);
-    // a smaller working set when this is set).
+    // mmap_flags |= MAP_POPULATE;
-    if (cfg_flags.sequential) {
-        posix_fadvise(file_descriptor, 0, 0, POSIX_FADV_SEQUENTIAL);
-    }
-    if (cfg_flags.populate) {
-        mmap_flags |= MAP_POPULATE;
-    }
 #endif
 
     struct stat sb;
@@ -265,27 +203,20 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bo
 
     size_t file_size = sb.st_size;
 
-    if (file_size == 0) {
+    void* mapped_data = mmap(NULL, file_size, PROT_READ, mmap_flags, file_descriptor, 0);
+
     close(file_descriptor);
-        return nullptr;
-    }
-
-    int mmap_prot = PROT_READ | (writable ? PROT_WRITE : 0);
-
-    void* mapped_data = mmap(nullptr, file_size, mmap_prot, mmap_flags, file_descriptor, 0);
 
     if (mapped_data == MAP_FAILED) {
-        close(file_descriptor);
         return nullptr;
     }
 
 #ifdef __linux__
-    if (cfg_flags.willneed) {
+    // performance flags used by llama.cpp
-        posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
+    // posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
-    }
 #endif
 
-    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size, file_descriptor);
+    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size);
 }
 
 #endif
@@ -564,6 +495,26 @@ sd_progress_cb_t sd_get_progress_callback() {
 void* sd_get_progress_callback_data() {
     return sd_progress_cb_data;
 }
+const char* sd_get_system_info() {
+    static char buffer[1024];
+    std::stringstream ss;
+    ss << "System Info: \n";
+    ss << "    SSE3 = " << ggml_cpu_has_sse3() << " | ";
+    ss << "    AVX = " << ggml_cpu_has_avx() << " | ";
+    ss << "    AVX2 = " << ggml_cpu_has_avx2() << " | ";
+    ss << "    AVX512 = " << ggml_cpu_has_avx512() << " | ";
+    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << " | ";
+    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << " | ";
+    ss << "    FMA = " << ggml_cpu_has_fma() << " | ";
+    ss << "    NEON = " << ggml_cpu_has_neon() << " | ";
+    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << " | ";
+    ss << "    F16C = " << ggml_cpu_has_f16c() << " | ";
+    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << " | ";
+    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << " | ";
+    ss << "    VSX = " << ggml_cpu_has_vsx() << " | ";
+    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
+    return buffer;
+}
 
 sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index) {
     const auto& shape = tensor.shape();
@@ -573,7 +524,17 @@ sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index)
     int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
     uint8_t* data = (uint8_t*)malloc(static_cast<size_t>(width * height * channel));
     GGML_ASSERT(data != nullptr);
-    preprocessing_tensor_frame_to_sd_image(tensor, frame_index, data);
+
+    for (int iw = 0; iw < width; ++iw) {
+        for (int ih = 0; ih < height; ++ih) {
+            for (int ic = 0; ic < channel; ++ic) {
+                float value                            = shape.size() == 5 ? tensor.index(iw, ih, frame_index, ic, 0)
+                                                                           : tensor.index(iw, ih, ic, frame_index);
+                value                                  = std::clamp(value, 0.0f, 1.0f);
+                data[(ih * width + iw) * channel + ic] = static_cast<uint8_t>(std::round(value * 255.0f));
+            }
+        }
+    }
     return {
         static_cast<uint32_t>(width),
         static_cast<uint32_t>(height),
@@ -757,100 +718,3 @@ std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::str
 
     return res;
 }
-
-// test if the backend is a specific one, e.g. "CUDA", "ROCm", "Vulkan" etc.
-bool sd_backend_is(ggml_backend_t backend, const std::string& name) {
-    if (!backend) {
-        return false;
-    }
-    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
-    if (!dev)
-        return false;
-    std::string dev_name = ggml_backend_dev_name(dev);
-    return dev_name.find(name) != std::string::npos;
-}
-
-ggml_backend_t sd_get_default_backend() {
-    ggml_backend_load_all_once();
-    static std::once_flag once;
-    std::call_once(once, []() {
-        size_t dev_count = ggml_backend_dev_count();
-        if (dev_count == 0) {
-            LOG_ERROR("No devices found!");
-        } else {
-            LOG_DEBUG("Found %zu backend devices:", dev_count);
-            for (size_t i = 0; i < dev_count; ++i) {
-                auto dev = ggml_backend_dev_get(i);
-                LOG_DEBUG("#%zu: %s", i, ggml_backend_dev_name(dev));
-            }
-        }
-    });
-    ggml_backend_t backend   = nullptr;
-    const char* SD_VK_DEVICE = getenv("SD_VK_DEVICE");
-    if (SD_VK_DEVICE != nullptr) {
-        std::string sd_vk_device_str = SD_VK_DEVICE;
-        try {
-            unsigned long long device  = std::stoull(sd_vk_device_str);
-            std::string vk_device_name = "Vulkan" + std::to_string(device);
-            if (backend_name_exists(vk_device_name)) {
-                LOG_INFO("Selecting %s as main device by env var SD_VK_DEVICE", vk_device_name.c_str());
-                backend = init_named_backend(vk_device_name);
-                if (!backend) {
-                    LOG_WARN("Device %s requested by SD_VK_DEVICE failed to init. Falling back to the default device.", vk_device_name.c_str());
-                }
-            } else {
-                LOG_WARN("Device %s requested by SD_VK_DEVICE was not found. Falling back to the default device.", vk_device_name.c_str());
-            }
-        } catch (const std::invalid_argument&) {
-            LOG_WARN("SD_VK_DEVICE environment variable is not a valid integer (%s). Falling back to the default device.", SD_VK_DEVICE);
-        } catch (const std::out_of_range&) {
-            LOG_WARN("SD_VK_DEVICE environment variable value is out of range for `unsigned long long` type (%s). Falling back to the default device.", SD_VK_DEVICE);
-        }
-    }
-
-    if (!backend) {
-        std::string dev_name = get_default_backend_name();
-        backend              = init_named_backend(dev_name);
-        if (!backend && !dev_name.empty()) {
-            LOG_WARN("device %s failed to init", dev_name.c_str());
-        }
-    }
-
-    if (!backend) {
-        LOG_WARN("loading CPU backend");
-        backend = ggml_backend_cpu_init();
-    }
-
-    if (ggml_backend_is_cpu(backend)) {
-        LOG_DEBUG("Using CPU backend");
-    }
-
-    return backend;
-}
-
-// namespace is needed to avoid conflicts with ggml_backend_extend.hpp
-namespace ggml_cpu {
-#include "ggml-cpu.h"
-}
-
-const char* sd_get_system_info() {
-    using namespace ggml_cpu;
-    static char buffer[1024];
-    std::stringstream ss;
-    ss << "System Info: \n";
-    ss << "    SSE3 = " << ggml_cpu_has_sse3() << " | ";
-    ss << "    AVX = " << ggml_cpu_has_avx() << " | ";
-    ss << "    AVX2 = " << ggml_cpu_has_avx2() << " | ";
-    ss << "    AVX512 = " << ggml_cpu_has_avx512() << " | ";
-    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << " | ";
-    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << " | ";
-    ss << "    FMA = " << ggml_cpu_has_fma() << " | ";
-    ss << "    NEON = " << ggml_cpu_has_neon() << " | ";
-    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << " | ";
-    ss << "    F16C = " << ggml_cpu_has_f16c() << " | ";
-    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << " | ";
-    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << " | ";
-    ss << "    VSX = " << ggml_cpu_has_vsx() << " | ";
-    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
-    return buffer;
-}

src/util.h

@@ -6,7 +6,6 @@
 #include <string>
 #include <vector>
 
-#include "ggml-backend.h"
 #include "stable-diffusion.h"
 #include "tensor.hpp"
 
@@ -42,7 +41,7 @@ sd::Tensor<float> clip_preprocess(const sd::Tensor<float>& image, int target_wid
 
 class MmapWrapper {
 public:
-    static std::unique_ptr<MmapWrapper> create(const std::string& filename, bool writable = false);
+    static std::unique_ptr<MmapWrapper> create(const std::string& filename);
 
     virtual ~MmapWrapper() = default;
 
@@ -52,7 +51,6 @@ public:
     MmapWrapper& operator=(MmapWrapper&&)      = delete;
 
     const uint8_t* data() const { return static_cast<uint8_t*>(data_); }
-    uint8_t* writable_data() { return static_cast<uint8_t*>(data_); }
     size_t size() const { return size_; }
     bool copy_data(void* buf, size_t n, size_t offset) const;
 
@@ -84,10 +82,6 @@ int sd_get_preview_interval();
 bool sd_should_preview_denoised();
 bool sd_should_preview_noisy();
 
-// test if the backend is a specific one, e.g. "CUDA", "ROCm", "Vulkan" etc.
-bool sd_backend_is(ggml_backend_t backend, const std::string& name);
-ggml_backend_t sd_get_default_backend();
-
 #define LOG_DEBUG(format, ...) log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #define LOG_INFO(format, ...) log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #define LOG_WARN(format, ...) log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)

src/vae.hpp

@@ -71,7 +71,7 @@ public:
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
-        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1) {
+        } else if (version == VERSION_CHROMA_RADIANCE) {
             scale_factor = 1;
         }
         return scale_factor;
@@ -142,9 +142,8 @@ public:
                                    "vae encode compute failed while processing a tile");
         } else {
             output = _compute(n_threads, input, false);
-        }
-
             free_compute_buffer();
+        }
 
         if (output.empty()) {
             LOG_ERROR("vae encode compute failed");

src/wan.hpp

@@ -692,7 +692,6 @@ namespace WAN {
             } else {
                 x = conv1->forward(ctx, x);
             }
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.prelude", "x");
 
             // downsamples
             std::vector<int64_t> dims = {dim};
@@ -718,14 +717,12 @@ namespace WAN {
                         x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
                     }
                 }
-                // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.down." + std::to_string(i), "x");
             }
 
             // middle
             x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
             x = middle_1->forward(ctx, x, b);
             x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.mid", "x");
 
             // head
             x = head_0->forward(ctx, x);
@@ -866,13 +863,11 @@ namespace WAN {
             } else {
                 x = conv1->forward(ctx, x);
             }
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.prelude", "x");
 
             // middle
             x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
             x = middle_1->forward(ctx, x, b);
             x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.mid", "x");
 
             // upsamples
             std::vector<int64_t> dims = {dim_mult[dim_mult.size() - 1] * dim};
@@ -898,7 +893,6 @@ namespace WAN {
                         x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
                     }
                 }
-                // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.up." + std::to_string(i), "x");
             }
 
             // head
@@ -1037,7 +1031,6 @@ namespace WAN {
             if (wan2_2) {
                 x = patchify(ctx->ggml_ctx, x, 2, b);
             }
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encode.prelude", "x");
 
             auto encoder = std::dynamic_pointer_cast<Encoder3d>(blocks["encoder"]);
             auto conv1   = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
@@ -1058,7 +1051,6 @@ namespace WAN {
             }
             out     = conv1->forward(ctx, out);
             auto mu = ggml_ext_chunk(ctx->ggml_ctx, out, 2, 3)[0];
-            // sd::ggml_graph_cut::mark_graph_cut(mu, "wan_vae.encode.final", "mu");
             clear_cache();
             return mu;
         }
@@ -1076,7 +1068,6 @@ namespace WAN {
 
             int64_t iter_ = z->ne[2];
             auto x        = conv2->forward(ctx, z);
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decode.prelude", "x");
             ggml_tensor* out;
             for (int i = 0; i < iter_; i++) {
                 _conv_idx = 0;
@@ -1092,7 +1083,6 @@ namespace WAN {
             if (wan2_2) {
                 out = unpatchify(ctx->ggml_ctx, out, 2, b);
             }
-            // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode.final", "out");
             clear_cache();
             return out;
         }
@@ -1108,14 +1098,12 @@ namespace WAN {
             auto conv2   = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv2"]);
 
             auto x    = conv2->forward(ctx, z);
-            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decode_partial.prelude", "x");
             auto in   = ggml_ext_slice(ctx->ggml_ctx, x, 2, i, i + 1);  // [b*c, 1, h, w]
             _conv_idx = 0;
             auto out  = decoder->forward(ctx, in, b, _feat_map, _conv_idx, i);
             if (wan2_2) {
                 out = unpatchify(ctx->ggml_ctx, out, 2, b);
             }
-            // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode_partial.final", "out");
             return out;
         }
     };
@@ -1996,13 +1984,6 @@ namespace WAN {
                 c = ggml_reshape_3d(ctx->ggml_ctx, c, c->ne[0] * c->ne[1] * c->ne[2], c->ne[3] / N, N);  // [N, dim, t_len*h_len*w_len]
                 c = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, c, 1, 0, 2, 3));  // [N, t_len*h_len*w_len, dim]
             }
-            sd::ggml_graph_cut::mark_graph_cut(x, "wan.prelude", "x");
-            // sd::ggml_graph_cut::mark_graph_cut(e, "wan.prelude", "e");
-            // sd::ggml_graph_cut::mark_graph_cut(e0, "wan.prelude", "e0");
-            // sd::ggml_graph_cut::mark_graph_cut(context, "wan.prelude", "context");
-            if (c != nullptr) {
-                sd::ggml_graph_cut::mark_graph_cut(c, "wan.prelude", "c");
-            }
 
             auto x_orig = x;
 
@@ -2023,10 +2004,6 @@ namespace WAN {
                     c_skip      = ggml_ext_scale(ctx->ggml_ctx, c_skip, vace_strength);
                     x           = ggml_add(ctx->ggml_ctx, x, c_skip);
                 }
-                sd::ggml_graph_cut::mark_graph_cut(x, "wan.blocks." + std::to_string(i), "x");
-                if (c != nullptr) {
-                    sd::ggml_graph_cut::mark_graph_cut(c, "wan.blocks." + std::to_string(i), "c");
-                }
             }
 
             x = head->forward(ctx, x, e);  // [N, t_len*h_len*w_len, pt*ph*pw*out_dim]

src/z_image.hpp

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