feat: move circular padding from context to per-generation params (#1748)

Co-authored-by: phil2sat <phil2sat@users.noreply.github.com>

.github/workflows/build.yml

@@ -15,7 +15,7 @@ on:
       [
         ".github/workflows/**",
         ".dockerignore",
-        "Dockerfile*",
+        "docker/Dockerfile*",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -32,7 +32,7 @@ on:
       [
         ".github/workflows/**",
         ".dockerignore",
-        "Dockerfile*",
+        "docker/Dockerfile*",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -264,7 +264,7 @@ jobs:
           context: .
           platforms: ${{ matrix.platform }}
           push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-          file: Dockerfile.${{ matrix.variant }}
+          file: docker/Dockerfile.${{ matrix.variant }}
           tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}${{ matrix.tag-suffix }}
           labels: ${{ steps.meta.outputs.labels }}
           annotations: ${{ steps.meta.outputs.annotations }}

CONTRIBUTING.md

@@ -27,7 +27,7 @@ PRs should include:
 
 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.
+Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `scripts/format-code.sh` or `scripts/format-code.ps1` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
 
 Naming conventions:
 

README.md

@@ -48,6 +48,7 @@ API and command-line option may change frequently.***
     - [PiD](./docs/pid.md)
     - [LongCat Image](./docs/longcat_image.md)
     - [Z-Image](./docs/z_image.md)
+    - [MiniT2I](./docs/minit2i.md)
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
@@ -179,19 +180,3 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=leejet/stable-diffusion.cpp&type=Date)](https://star-history.com/#leejet/stable-diffusion.cpp&Date)
-
-## References
-
-- [ggml](https://github.com/ggml-org/ggml)
-- [diffusers](https://github.com/huggingface/diffusers)
-- [stable-diffusion](https://github.com/CompVis/stable-diffusion)
-- [sd3-ref](https://github.com/Stability-AI/sd3-ref)
-- [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
-- [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui)
-- [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
-- [k-diffusion](https://github.com/crowsonkb/k-diffusion)
-- [latent-consistency-model](https://github.com/luosiallen/latent-consistency-model)
-- [generative-models](https://github.com/Stability-AI/generative-models/)
-- [PhotoMaker](https://github.com/TencentARC/PhotoMaker)
-- [Wan2.1](https://github.com/Wan-Video/Wan2.1)
-- [Wan2.2](https://github.com/Wan-Video/Wan2.2)

docs/anima.md

@@ -15,7 +15,7 @@
 ## Examples
 
 ```sh
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\anima-preview.safetensors --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="anima image example" src="../assets/anima/example.png" />

docs/backend.md

@@ -51,6 +51,97 @@ Module names are case-insensitive. Hyphens and underscores in module names are i
 sd-cli -m model.safetensors -p "a cat" --backend all=cuda0,te=cpu
 ```
 
+## Multiple devices per module (layer split)
+
+A `--backend` module assignment can list several devices separated by `&`:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1"
+```
+
+The module's transformer blocks are then distributed across the listed devices
+in contiguous ranges sized proportionally to each device's free memory (minus a
+compute-buffer headroom of about 2 GiB per device), and the
+module's graphs are executed with a `ggml_backend_sched` that runs each block
+on the device holding its weights, copying the residual stream at the range
+boundaries. The first device in the list is the module's main device: it also
+holds the non-block tensors (embeddings, final norms, small sub-runners such as
+CLIP models or projectors) and the graph inputs/outputs.
+
+Layer split is supported for the `diffusion` and `te` modules. For `te` it
+applies to the dominant text encoder (`t5xxl` or the LLM); other modules accept
+only a single device. If the module has no recognizable transformer blocks, the
+assignment falls back to the first listed device.
+
+`--params-backend` accepts no device lists. If the module has no explicit
+params assignment, each block range's parameters are loaded directly to (and,
+with `--params-backend diffusion=disk`, released directly from) its own device;
+an explicit assignment such as `te=cpu` keeps the parameters on that backend
+and stages each range to its device on demand.
+
+Layer split cannot be combined with `--max-vram` graph-cut segmentation or
+`--stream-layers` for the split module; those are single-device mechanisms and
+are disabled for it.
+
+Use `--list-devices` to see the device names available on the system.
+
+### Row split (`--split-mode row`)
+
+`--split-mode` selects how a multi-device module distributes its weights:
+`layer` (the default, described above) or `row`. It accepts a single mode or
+per-module assignments:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1" --split-mode row
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1,te=cuda0&cuda1" --split-mode diffusion=row,te=layer
+```
+
+In row mode the module keeps executing on its main (first listed) device, but
+its transformer-block matmul weights are allocated in the backend's row-split
+buffer type, which slices each weight's rows across the listed devices in
+proportion to free memory and runs those matmuls on all devices in parallel.
+Compared to a layer split this uses all GPUs within every layer (instead of
+sequentially device by device) at the cost of a cross-device reduction per
+matmul - usually the faster option when the devices have fast interconnect.
+
+Row split requires backend support for split buffers and is currently
+available on CUDA only; on other backends (or when the listed devices belong
+to different backend registries) the module falls back to a layer split.
+Embeddings, normalization weights, biases and other non-block tensors stay in
+regular buffers on the main device.
+
+Direct ("immediately") LoRA application cannot patch row-split tensors; with
+`--split-mode row` the automatic LoRA mode selects runtime application, and an
+explicit `--lora-apply-mode immediately` skips the split tensors with a
+warning.
+
+## Automatic placement (`--auto-fit`)
+
+`--auto-fit` derives the `diffusion` / `te` / `vae` placements from the model
+metadata and the per-device memory budgets, then feeds them into the same
+backend assignment mechanism described above (the chosen specs are printed).
+`--backend` and `--params-backend` are ignored while auto-fit is enabled.
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --auto-fit
+sd-cli -m model.safetensors -p "a cat" --auto-fit --max-vram cuda0=8,cuda1=14
+sd-cli -m model.safetensors -p "a cat" --auto-fit --split-mode row
+```
+
+Budgets reuse `--max-vram`: a positive per-device value caps what auto-fit
+plans with on that device, a negative value means "free memory minus that many
+GiB", and with no budget set each device's free memory minus a 512 MiB margin
+is used. (The same values still drive graph-cut segmented execution for
+modules that end up on a single device.)
+
+When everything fits resident, components are simply spread across the
+available GPUs. When it does not, auto-fit switches to time-share mode: the
+heavy components get `disk` params residency (loaded for their phase, freed
+after), and a component too large for any single device is split across all
+GPUs with the layer/row split mechanism (`--split-mode` selects which, layer
+by default). Components that fit nowhere fall back to the CPU. If a VAE decode
+still runs out of memory, tiling is enabled and the decode retried once.
+
 ## Modules
 
 | Module | Purpose | Accepted names |

docs/boogu_image.md

@@ -17,7 +17,7 @@ Boogu Image uses a Boogu diffusion transformer, the FLUX VAE, and Qwen3-VL as th
 ### Boogu Image Base
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\boogu_image_base_bf16.safetensors --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ae.sft -p "a lovely cat" --diffusion-fa -v --offload-to-cpu
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\boogu_image_base_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\models\vae\ae.sft -p "a lovely cat" --diffusion-fa -v --offload-to-cpu
 ```
 
 <img width="256" alt="Boogu Image Base example" src="../assets/boogu/example.png" />
@@ -25,7 +25,7 @@ Boogu Image uses a Boogu diffusion transformer, the FLUX VAE, and Qwen3-VL as th
 ### Boogu Image Edit
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\boogu_image_edit_bf16.safetensors --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --llm_vision ..\..\llm\mmproj-Qwen3VL-8B-Instruct-F16.gguf --vae ..\..\ComfyUI\models\vae\ae.sft --diffusion-fa -v --offload-to-cpu -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'boogu.cpp'"
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\boogu_image_edit_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\mmproj-Qwen3VL-8B-Instruct-F16.gguf --vae ..\models\vae\ae.sft --diffusion-fa -v --offload-to-cpu -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'boogu.cpp'"
 ```
 
 <img width="256" alt="Boogu Image Edit example" src="../assets/boogu/edit_example.png" />

docs/chroma.md

@@ -15,7 +15,7 @@ You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4G
 You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF), this way you don't have to do the conversion yourself.
 
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
 ```
 
 ## Run

docs/chroma_radiance.md

@@ -12,7 +12,7 @@
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\..\ComfyUI\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\Chroma1-Radiance-v0.4-Q8_0.gguf --t5xxl ..\models\clip\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma  radiance cpp'" --cfg-scale 4.0 --sampling-method euler -v
 ```
 
 <img alt="Chroma1-Radiance" src="../assets/flux/chroma1-radiance.png" />

docs/docker.md

@@ -19,7 +19,7 @@ docker run --rm --init -v /path/to/models:/models -v /path/to/output/:/output -p
 ## Building using Docker
 
 ```shell
-docker build -t sd .
+docker build -f docker/Dockerfile -t sd .
 ```
 
 ## Building variants using Docker
@@ -27,7 +27,7 @@ docker build -t sd .
 Vulkan:
 
 ```shell
-docker build -f Dockerfile.vulkan -t sd .
+docker build -f docker/Dockerfile.vulkan -t sd .
 ```
 
 ## Run locally built image's CLI

docs/ernie_image.md

@@ -21,7 +21,7 @@ You can run ERNIE-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — o
 ### ERNIE-Image-Turbo
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-turbo.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 8 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\ernie-image-turbo.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 8 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img width="256" alt="ERNIE-Image Turbo example" src="../assets/ernie_image/turbo_example.png" />
@@ -29,7 +29,7 @@ You can run ERNIE-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — o
 ### ERNIE-Image
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\ernie-image-UD-Q4_K_M.gguf --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img width="256" alt="ERNIE-Image example" src="../assets/ernie_image/example.png" />

docs/flux.md

@@ -17,7 +17,7 @@ You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://hu
 
 For example:
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
 ```
 
 ## Run

docs/flux2.md

@@ -16,7 +16,7 @@
 ### Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
 ```
 
 <img alt="flux2 example" src="../assets/flux2/example.png" />
@@ -42,19 +42,19 @@
 ### Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-4b" src="../assets/flux2/flux2-klein-4b.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
 ```
 
 <img alt="flux2-klein-4b-edit" src="../assets/flux2/flux2-klein-4b-edit.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-base-4b" src="../assets/flux2/flux2-klein-base-4b.png" />
@@ -78,19 +78,19 @@
 ### Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-9b" src="../assets/flux2/flux2-klein-9b.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
 ```
 
 <img alt="flux2-klein-9b-edit" src="../assets/flux2/flux2-klein-9b-edit.png" />
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\models\vae\flux2_ae.safetensors  --llm ..\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="flux2-klein-base-9b" src="../assets/flux2/flux2-klein-base-9b.png" />

docs/hidream_o1_image.md

@@ -12,7 +12,7 @@
 ### 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 
+.\bin\Release\sd-cli.exe -m  ..\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
 ```
 

docs/ideogram4.md

@@ -34,7 +34,7 @@ If you want lower VRAM usage, you can change the quantization from q8_0 to a low
 ## Examples
 
 ```sh
-.\bin\Release\sd-cli.exe --diffusion-model ideogram4-Q8_0.gguf --uncond-diffusion-model ideogram4_uncond-Q8_0.gguf --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors -p '{"high_level_description":"A square 1024 x 1024 luxury fashion magazine cover featuring exactly one short chubby fluffy cat as the main model. The cat sits on a soft ivory studio floor, facing the viewer with a stylish calm expression, wearing tiny black sunglasses, a red silk scarf, and a small gold collar charm. In front of the cat on the floor is a wide horizontal luxury nameplate that clearly reads ideogram4.cpp. The whole design feels premium, fashionable, clean, and editorial.","style_description":{"aesthetics":"luxury fashion magazine cover, high-end pet couture campaign, minimalist editorial design, elegant studio photography, soft paper texture, refined typography, fashionable and polished","lighting":"Soft diffused studio lighting, gentle spotlight on the cat, subtle floor shadow, warm ivory highlights, clean separation between subject and background","photo":"high-resolution fashion editorial photography look, front-facing cat portrait, crisp fur details, glossy sunglasses, clear readable nameplate text, shallow depth of field","medium":"mixed media fashion photography and premium editorial graphic design","color_palette":["#F4EFE7","#111111","#D8B56D","#B73A3A","#FFFFFF","#8A7A6A"]},"compositional_deconstruction":{"canvas":"Square 1024 x 1024 canvas with a normal upright orientation. Do not rotate the poster or any text. Use a clean fashion magazine cover layout.","background":"Warm ivory studio backdrop with subtle paper grain, a soft spotlight gradient, faint floor shadow, and a few minimal gold editorial lines. The background is spacious, premium, and uncluttered.","layout":"Top center has a small elegant headline. Center area features one cat as the main fashion model. Lower foreground has a wide horizontal luxury nameplate placed on the floor in front of the cat. Bottom center has a small footer. All text is horizontal, upright, and readable left to right.","elements":[{"type":"text","desc":"Top center headline reading LOOK WHAT I FOUND in a refined high-fashion serif font. The headline is horizontal, centered, elegant, and secondary to the nameplate text."},{"type":"obj","desc":"Exactly one short chubby fluffy cat sitting in the center like a luxury fashion model. The cat has a large round head, compact body, short legs, soft detailed fur, expressive eyes, and a calm confident pose. The cat is cute and rounded, not tall, not stretched, not duplicated."},{"type":"obj","desc":"Tiny glossy black sunglasses worn naturally by the cat, slightly oversized but still showing the cat face clearly. The sunglasses add a chic fashion-editorial attitude."},{"type":"obj","desc":"A red silk scarf tied neatly around the cat neck, with soft folds and a couture feeling. The scarf must not cover the cat face or the nameplate."},{"type":"obj","desc":"A small gold collar charm or fashion accessory under the scarf, subtle and premium, adding a luxury campaign detail."},{"type":"obj","desc":"In the lower foreground, place a wide horizontal luxury nameplate on the floor in front of the cat. The nameplate is low, flat, landscape-oriented, much wider than tall, like a fashion show seat card or premium display plaque. It is centered, front-facing, level, and fully visible. It must not become vertical, tall, standing, rotated, or side-facing."},{"type":"text","desc":"Print the exact text ideogram4.cpp only on the wide horizontal nameplate. Use clean bold black lettering, perfectly spelled, lowercase, with the number 4 and .cpp extension. The text must fit completely inside the nameplate, stay horizontal, and be readable from left to right."},{"type":"obj","desc":"Add sparse premium editorial accents around the edges: thin gold lines, small code brackets, tiny cursor marks, subtle dots, and minimal geometric details. No extra cats, no stickers, no animal faces, no busy decorations."},{"type":"text","desc":"Bottom center footer reading tiny paws, big compile energy in a small refined monospace or editorial font. The footer is horizontal, centered, understated, and much smaller than the nameplate text."}]}}'  --diffusion-fa -v --offload-to-cpu -H 1024 -W 1024
+.\bin\Release\sd-cli.exe --diffusion-model ideogram4-Q8_0.gguf --uncond-diffusion-model ideogram4_uncond-Q8_0.gguf --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\models\vae\flux2_ae.safetensors -p '{"high_level_description":"A square 1024 x 1024 luxury fashion magazine cover featuring exactly one short chubby fluffy cat as the main model. The cat sits on a soft ivory studio floor, facing the viewer with a stylish calm expression, wearing tiny black sunglasses, a red silk scarf, and a small gold collar charm. In front of the cat on the floor is a wide horizontal luxury nameplate that clearly reads ideogram4.cpp. The whole design feels premium, fashionable, clean, and editorial.","style_description":{"aesthetics":"luxury fashion magazine cover, high-end pet couture campaign, minimalist editorial design, elegant studio photography, soft paper texture, refined typography, fashionable and polished","lighting":"Soft diffused studio lighting, gentle spotlight on the cat, subtle floor shadow, warm ivory highlights, clean separation between subject and background","photo":"high-resolution fashion editorial photography look, front-facing cat portrait, crisp fur details, glossy sunglasses, clear readable nameplate text, shallow depth of field","medium":"mixed media fashion photography and premium editorial graphic design","color_palette":["#F4EFE7","#111111","#D8B56D","#B73A3A","#FFFFFF","#8A7A6A"]},"compositional_deconstruction":{"canvas":"Square 1024 x 1024 canvas with a normal upright orientation. Do not rotate the poster or any text. Use a clean fashion magazine cover layout.","background":"Warm ivory studio backdrop with subtle paper grain, a soft spotlight gradient, faint floor shadow, and a few minimal gold editorial lines. The background is spacious, premium, and uncluttered.","layout":"Top center has a small elegant headline. Center area features one cat as the main fashion model. Lower foreground has a wide horizontal luxury nameplate placed on the floor in front of the cat. Bottom center has a small footer. All text is horizontal, upright, and readable left to right.","elements":[{"type":"text","desc":"Top center headline reading LOOK WHAT I FOUND in a refined high-fashion serif font. The headline is horizontal, centered, elegant, and secondary to the nameplate text."},{"type":"obj","desc":"Exactly one short chubby fluffy cat sitting in the center like a luxury fashion model. The cat has a large round head, compact body, short legs, soft detailed fur, expressive eyes, and a calm confident pose. The cat is cute and rounded, not tall, not stretched, not duplicated."},{"type":"obj","desc":"Tiny glossy black sunglasses worn naturally by the cat, slightly oversized but still showing the cat face clearly. The sunglasses add a chic fashion-editorial attitude."},{"type":"obj","desc":"A red silk scarf tied neatly around the cat neck, with soft folds and a couture feeling. The scarf must not cover the cat face or the nameplate."},{"type":"obj","desc":"A small gold collar charm or fashion accessory under the scarf, subtle and premium, adding a luxury campaign detail."},{"type":"obj","desc":"In the lower foreground, place a wide horizontal luxury nameplate on the floor in front of the cat. The nameplate is low, flat, landscape-oriented, much wider than tall, like a fashion show seat card or premium display plaque. It is centered, front-facing, level, and fully visible. It must not become vertical, tall, standing, rotated, or side-facing."},{"type":"text","desc":"Print the exact text ideogram4.cpp only on the wide horizontal nameplate. Use clean bold black lettering, perfectly spelled, lowercase, with the number 4 and .cpp extension. The text must fit completely inside the nameplate, stay horizontal, and be readable from left to right."},{"type":"obj","desc":"Add sparse premium editorial accents around the edges: thin gold lines, small code brackets, tiny cursor marks, subtle dots, and minimal geometric details. No extra cats, no stickers, no animal faces, no busy decorations."},{"type":"text","desc":"Bottom center footer reading tiny paws, big compile energy in a small refined monospace or editorial font. The footer is horizontal, centered, understated, and much smaller than the nameplate text."}]}}'  --diffusion-fa -v --offload-to-cpu -H 1024 -W 1024
 ```
 
 <img alt="ideogram4 image example" src="../assets/ideogram4/example.png" />

docs/imatrix.md

@@ -0,0 +1,59 @@
+# Importance Matrix (imatrix) Quantization
+
+## What is an Importance Matrix?
+
+Quantization reduces the precision of a model's weights, decreasing its size and computational requirements. However, this can lead to a loss of quality. An importance matrix helps mitigate this by identifying which weights are *most* important for the model's performance. During quantization, these important weights are preserved with higher precision, while less important weights are quantized more aggressively.  This allows for better overall quality at a given quantization level.
+
+This originates from work done with language models in [llama.cpp](https://github.com/ggml-org/llama.cpp/blob/master/tools/imatrix/README.md).
+
+## Usage
+
+The imatrix feature involves two main steps: *training* the matrix and *using* it during quantization.
+
+### Training the Importance Matrix
+
+To generate an imatrix, run stable-diffusion.cpp with the `--imat-out` flag, specifying the output filename.  This process runs alongside normal image generation.
+
+```bash
+sd.exe [same exact parameters as normal generation] --imat-out imatrix.dat
+```
+
+*   **`[same exact parameters as normal generation]`**:  Use the same command-line arguments you would normally use for image generation (e.g., prompt, dimensions, sampling method, etc.).
+*   **`--imat-out imatrix.dat`**: Specifies the output file for the generated imatrix.
+
+You can generate multiple images at once using the `-b` flag to speed up the training process.
+
+### Continuing Training an Existing Matrix
+
+If you want to refine an existing imatrix, use the `--imat-in` flag *in addition* to `--imat-out`. This will load the existing matrix and continue training it.
+
+```bash
+sd.exe [same exact parameters as normal generation] --imat-out imatrix.dat --imat-in imatrix.dat
+```
+With that, you can train and refine the imatrix while generating images like you'd normally do. 
+
+### Using Multiple Matrices
+
+You can load and merge multiple imatrices together:
+
+```bash
+sd.exe [same exact parameters as normal generation] --imat-out imatrix.dat --imat-in imatrix.dat --imat-in imatrix2.dat
+```
+
+### Quantizing with an Importance Matrix
+
+To quantize a model using a trained imatrix, use the `-M convert` option (or equivalent quantization command) and the `--imat-in` flag, specifying the imatrix file.
+
+```bash
+sd.exe -M convert [same exact parameters as normal quantization] --imat-in imatrix.dat
+```
+
+*   **`[same exact parameters as normal quantization]`**: Use the same command-line arguments you would normally use for quantization (e.g., target quantization method, input/output filenames).
+*   **`--imat-in imatrix.dat`**: Specifies the imatrix file to use during quantization.  You can specify multiple `--imat-in` flags to combine multiple matrices.
+
+## Important Considerations
+
+*   The quality of the imatrix depends on the prompts and settings used during training. Use prompts and settings representative of the types of images you intend to generate for the best results.
+*   Experiment with different training parameters (e.g., number of images, prompt variations) to optimize the imatrix for your specific use case.
+*   The performance impact of training an imatrix during image generation or using an imatrix for quantization is negligible.
+*   Using already quantized models to train the imatrix seems to be working fine.

docs/kontext.md

@@ -16,7 +16,7 @@ You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4
 You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF), this way you don't have to do the conversion yourself.
 
 ```
-.\bin\Release\sd-cli.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
+.\bin\Release\sd-cli.exe -M convert -m ..\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
 ```
 
 ## Run

docs/krea2.md

@@ -21,7 +21,7 @@ Krea2 uses a Krea2 diffusion transformer, the Wan2.1 VAE, and Qwen3-VL 4B as the
 ### Krea2
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\Krea-2-Raw-Q8_0.gguf --llm ..\..\ComfyUI\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors -p "a lovely cat holding a sign says 'krea2.cpp'"  --diffusion-fa -v --offload-to-cpu
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\Krea-2-Raw-Q8_0.gguf --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf --vae ..\models\vae\wan_2.1_vae.safetensors -p "a lovely cat holding a sign says 'krea2.cpp'"  --diffusion-fa -v --offload-to-cpu
 ```
 
 <img width="256" alt="Krea2 Raw example" src="../assets/krea2/example.png" />

docs/lens.md

@@ -18,7 +18,7 @@ Lens uses a Lens diffusion transformer, the FLUX.2 VAE, and GPT-OSS-20B as the L
 ### Lens
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 5.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\lens_bf16.safetensors --llm "..\models\text_encoders\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\models\vae\flux2_ae.safetensors --cfg-scale 5.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v
 ```
 
 <img width="256" alt="Lens example" src="../assets/lens/example.png" />
@@ -26,7 +26,7 @@ Lens uses a Lens diffusion transformer, the FLUX.2 VAE, and GPT-OSS-20B as the L
 ### Lens Turbo
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_turbo_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 1.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v --steps 4
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\lens_turbo_bf16.safetensors --llm "..\models\text_encoders\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\models\vae\flux2_ae.safetensors --cfg-scale 1.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v --steps 4
 ```
 
 <img width="256" alt="Lens Turbo example" src="../assets/lens/turbo_example.png" />

docs/longcat_image.md

@@ -24,7 +24,7 @@ LongCat uses quoted text for character-level text rendering. Put target text ins
 ### LongCat Image
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\LongCat-Image-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ae.sft --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf -p "a lovely cat holding a sign says 'longcat.cpp'" --cfg-scale 5.0 --sampling-method euler --flow-shift 3 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\LongCat-Image-Q4_K_M.gguf --vae ..\models\vae\ae.sft --llm ..\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf -p "a lovely cat holding a sign says 'longcat.cpp'" --cfg-scale 5.0 --sampling-method euler --flow-shift 3 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="longcat example" src="../assets/longcat/example.png" />

docs/ltx2.md

@@ -21,7 +21,7 @@
 ### LTX-2.3 dev T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "worst quality, low quality, blurry, distorted, artifacts" -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --fps 24 -o t2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "worst quality, low quality, blurry, distorted, artifacts" -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --fps 24 -o t2v.webm
 ```
 
 <video
@@ -33,7 +33,7 @@
 ### LTX-2.3 dev I2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o i2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o i2v.webm
 ```
 
 <video
@@ -45,7 +45,7 @@
 ### LTX-2.3 dev FLF2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png -o flf2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors  -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v  -W 1280 -H 720 --diffusion-fa --offload-to-cpu --video-frames 33 --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png -o flf2v.webm
 ```
 
 <video
@@ -61,7 +61,7 @@ LTX spatial latent upscale runs a model-backed x2 latent upsampler between the l
 Put `ltx-2.3-spatial-upscaler-x2-1.1.safetensors` under the directory passed to `--hires-upscalers-dir`, then use the model name without path or extension in `--hires-upscaler`.
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors --hires-upscalers-dir ..\..\ComfyUI\models\latent_upscale_models --hires-upscaler ltx-2.3-spatial-upscaler-x2-1.1 --hires --hires-steps 4 -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 640 -H 360 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o hires_i2v.webm
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors --hires-upscalers-dir ..\models\latent_upscale_models --hires-upscaler ltx-2.3-spatial-upscaler-x2-1.1 --hires --hires-steps 4 -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 640 -H 360 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o hires_i2v.webm
 ```
 
 By default, the hires refine pass uses the main sampler and scheduler, then trims the second-pass sigma schedule by `--hires-denoising-strength` (`0.7` by default). To reproduce a ComfyUI-style explicit refine schedule, pass custom hires sigmas:

docs/minit2i.md

@@ -0,0 +1,48 @@
+# How to Use
+
+MiniT2I uses a MiniT2I diffusion transformer and `google/flan-t5-large` as the text encoder.
+
+## Download weights
+
+- Download MiniT2I diffusion model
+    - safetensors: https://huggingface.co/MiniT2I/MiniT2I/tree/main/minit2i-b-16/transformer (`diffusion_pytorch_model.safetensors`)
+- Download flan-t5-large text encoder
+    - safetensors: https://huggingface.co/google/flan-t5-large/tree/main (`model.safetensors`)
+
+## Examples
+
+### Mac Metal
+
+```
+./bin/sd-cli \
+  --backend metal \
+  --diffusion-model ../models/minit2i/diffusion_pytorch_model.safetensors \
+  --t5xxl ../models/flan-t5-large/model.safetensors \
+  --prompt "a cat" \
+  --steps 100 \
+  --cfg-scale 6 \
+  --width 512 \
+  --height 512 \
+  --seed 42 \
+  --sampling-method euler \
+  --rng cpu \
+  --output minit2i_metal.png \
+  --threads 8
+```
+
+### CUDA with diffusion flash attention
+
+```
+./bin/sd-cli \
+  --diffusion-model ../models/minit2i/diffusion_pytorch_model.safetensors \
+  --t5xxl ../models/flan-t5-large/model.safetensors \
+  --prompt "a cat" \
+  --steps 100 \
+  --cfg-scale 6 \
+  --width 512 \
+  --height 512 \
+  --seed 42 \
+  --sampling-method euler \
+  --diffusion-fa \
+  --output minit2i_cuda.png
+```

docs/ovis_image.md

@@ -13,7 +13,7 @@
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ovis_image-Q4_0.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\ovis_2.5.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
+.\bin\Release\sd-cli.exe --diffusion-model  ovis_image-Q4_0.gguf --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\ovis_2.5.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
 ```
 
 <img alt="ovis image example" src="../assets/ovis_image/example.png" />

docs/pid.md

@@ -22,7 +22,7 @@ The official PiD model card should be checked before use. At the time of the ini
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\pid_flux1_512_to_2048_4step_bf16.safetensors --llm "..\..\ComfyUI\models\text_encoders\gemma_2_2b_it_elm_bf16.safetensors" --vae ..\..\ComfyUI\models\vae\ae.sft --vae-format flux --cfg-scale 1.0  -p "a lovely cat" -r ..\assets\ernie_image\turbo_example.png --diffusion-fa -v --steps 4 -H 2048 -W 2048 --rng cpu
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\pid_flux1_512_to_2048_4step_bf16.safetensors --llm "..\models\text_encoders\gemma_2_2b_it_elm_bf16.safetensors" --vae ..\models\vae\ae.sft --vae-format flux --cfg-scale 1.0  -p "a lovely cat" -r ..\assets\ernie_image\turbo_example.png --diffusion-fa -v --steps 4 -H 2048 -W 2048 --rng cpu
 ```
 
 Before:

docs/pulid.md

@@ -52,7 +52,7 @@ to a `.pulidembd` binary file (about 131 KB). Run it once per source
 person; the same file is reused for any number of generations.
 
 A reference Python script is provided alongside this docs file at
-[`script/pulid_extract_id.py`](../script/pulid_extract_id.py). It
+[`scripts/pulid_extract_id.py`](../scripts/pulid_extract_id.py). It
 requires:
 - A working CUDA / CPU PyTorch stack
 - `insightface`, `facexlib`, `eva-clip`, `torchvision`, `opencv-python`,

docs/qwen_image.md

@@ -14,7 +14,7 @@
 ## Examples
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
 ```
 
 <img alt="qwen example" src="../assets/qwen/example.png" />

docs/qwen_image_edit.md

@@ -23,7 +23,7 @@
 ### Qwen Image Edit
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
 ```
 
 <img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
@@ -32,7 +32,7 @@
 ### Qwen Image Edit 2509
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
 ```
 
 <img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />
@@ -42,7 +42,7 @@
 To use the new Qwen Image Edit 2511 mode, the  `--qwen-image-zero-cond-t` flag must be enabled; otherwise, image editing quality will degrade significantly.
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-edit-2511-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'"  --qwen-image-zero-cond-t
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\qwen-image-edit-2511-Q4_K_M.gguf --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'"  --qwen-image-zero-cond-t
 ```
 
 <img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2511.png" />

docs/sefi_image.md

@@ -13,8 +13,8 @@ The SeFi-Image family ships in three scales (1B / 2B / 5B) and three families (B
 Convert the transformer and text encoder to sd.cpp safetensors:
 
 ```bash
-python3 script/convert_sefi.py     <hf_repo_dir>                          <out_dir>/sefi_<scale>_<family>.safetensors
-python3 script/convert_qwen3_vl.py  <hf_repo_dir>/Qwen3-VL-XB-Instruct    <out_dir>/qwen3_vl_<X>b.safetensors
+python3 scripts/convert_sefi.py     <hf_repo_dir>                          <out_dir>/sefi_<scale>_<family>.safetensors
+python3 scripts/convert_qwen3_vl.py  <hf_repo_dir>/Qwen3-VL-XB-Instruct    <out_dir>/qwen3_vl_<X>b.safetensors
 ```
 
 ## Variant defaults

docs/wan.md

@@ -55,7 +55,7 @@
 ### Wan2.1 T2V 1.3B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -63,7 +63,7 @@
 ### Wan2.1 T2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.1_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -73,7 +73,7 @@
 ### Wan2.1 I2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.1_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -81,7 +81,7 @@
 ### Wan2.2 T2V A14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -89,7 +89,7 @@
 ### Wan2.2 I2V A14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -97,7 +97,7 @@
 ### Wan2.2 T2V A14B T2I
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
 ```
 
 <img width="832" height="480" alt="Wan2 2_14B_t2i" src="../assets/wan/Wan2.2_14B_t2i.png" />
@@ -105,7 +105,7 @@
 ### Wan2.2 T2V 14B with Lora
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\models\loras --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_t2v_lora.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -117,7 +117,7 @@
 #### T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\models\vae\wan2.2_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_5B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -125,7 +125,7 @@
 #### I2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\models\vae\wan2.2_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_5B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -133,7 +133,7 @@
 ### Wan2.1 FLF2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```
 
 
@@ -142,7 +142,7 @@
 ### Wan2.2 FLF2V 14B
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
 ```
 
 <video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -152,7 +152,7 @@
 #### T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -161,7 +161,7 @@
 #### R2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -172,7 +172,7 @@
 ```
 mkdir post+depth
 ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -182,7 +182,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### T2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -191,7 +191,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### R2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
@@ -201,7 +201,7 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 #### V2V
 
 ```
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>

docs/z_image.md

@@ -21,7 +21,7 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 ### Z-Image-Turbo
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512 --steps 8
+.\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512 --steps 8
 ```
 
 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
@@ -29,7 +29,7 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 ### Z-Image-Base
 
 ```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\z_image_bf16.safetensors --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\z_image_bf16.safetensors --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
 ```
 
 <img width="256" alt="z-image example" src="../assets/z_image/base_bf16.png" />

examples/cli/main.cpp

@@ -1,6 +1,7 @@
 #include <stdio.h>
 #include <string.h>
 #include <time.h>
+#include <algorithm>
 #include <cctype>
 #include <filesystem>
 #include <functional>
@@ -53,6 +54,9 @@ struct SDCliParams {
     bool metadata_brief      = false;
     bool metadata_all        = false;
 
+    std::string imatrix_out;
+    std::vector<std::string> imatrix_in;
+
     bool normal_exit = false;
 
     ArgOptions get_options() {
@@ -79,6 +83,11 @@ struct SDCliParams {
              "path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp",
              0,
              &preview_path},
+            {"",
+             "--imat-out",
+             "compute the imatrix for this run and save it to the provided path",
+             0,
+             &imatrix_out},
         };
 
         options.int_options = {
@@ -179,6 +188,14 @@ struct SDCliParams {
             return -1;
         };
 
+        auto on_imatrix_in_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
+            }
+            imatrix_in.push_back(argv[index]);
+            return 1;
+        };
+
         options.manual_options = {
             {"-M",
              "--mode",
@@ -192,6 +209,10 @@ struct SDCliParams {
              "--help",
              "show this help message and exit",
              on_help_arg},
+            {"",
+             "--imat-in",
+             "load an imatrix file for quantization or continued collection; can be specified multiple times",
+             on_imatrix_in_arg},
         };
 
         return options;
@@ -253,6 +274,7 @@ struct SDCliParams {
             << "  preview_fps: " << preview_fps << ",\n"
             << "  taesd_preview: " << (taesd_preview ? "true" : "false") << ",\n"
             << "  preview_noisy: " << (preview_noisy ? "true" : "false") << ",\n"
+            << "  imatrix_out: \"" << imatrix_out << "\",\n"
             << "  metadata_raw: " << (metadata_raw ? "true" : "false") << ",\n"
             << "  metadata_brief: " << (metadata_brief ? "true" : "false") << ",\n"
             << "  metadata_all: " << (metadata_all ? "true" : "false") << "\n"
@@ -459,7 +481,8 @@ 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;
+        int images_per_batch        = gen_params.batch_count > 0 ? std::max(1, num_results / gen_params.batch_count) : 1;
+        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx / images_per_batch;
         std::string params          = gen_params.embed_image_metadata
                                           ? get_image_params(ctx_params, gen_params, metadata_seed, cli_params.mode)
                                           : "";
@@ -605,13 +628,33 @@ int main(int argc, const char* argv[]) {
     LOG_DEBUG("%s", ctx_params.to_string().c_str());
     LOG_DEBUG("%s", gen_params.to_string().c_str());
 
+    if (!cli_params.imatrix_out.empty()) {
+        if (fs::exists(cli_params.imatrix_out) &&
+            std::find(cli_params.imatrix_in.begin(), cli_params.imatrix_in.end(), cli_params.imatrix_out) == cli_params.imatrix_in.end()) {
+            LOG_WARN("imatrix file '%s' already exists and will be overwritten", cli_params.imatrix_out.c_str());
+        }
+        enable_imatrix_collection();
+    }
+
+    for (const auto& in_file : cli_params.imatrix_in) {
+        LOG_INFO("loading imatrix from '%s'", in_file.c_str());
+        if (!load_imatrix(in_file.c_str())) {
+            LOG_WARN("failed to load imatrix from '%s'", in_file.c_str());
+        }
+    }
+
     if (cli_params.mode == CONVERT) {
-        bool success = convert(ctx_params.model_path.c_str(),
-                               ctx_params.vae_path.c_str(),
-                               cli_params.output_path.c_str(),
-                               ctx_params.wtype,
-                               ctx_params.tensor_type_rules.c_str(),
-                               cli_params.convert_name);
+        bool success = convert_with_components(ctx_params.model_path.c_str(),
+                                               ctx_params.clip_l_path.c_str(),
+                                               ctx_params.clip_g_path.c_str(),
+                                               ctx_params.t5xxl_path.c_str(),
+                                               ctx_params.diffusion_model_path.c_str(),
+                                               ctx_params.vae_path.c_str(),
+                                               cli_params.output_path.c_str(),
+                                               ctx_params.wtype,
+                                               ctx_params.tensor_type_rules.c_str(),
+                                               cli_params.convert_name,
+                                               ctx_params.n_threads);
         if (!success) {
             LOG_ERROR("convert '%s'/'%s' to '%s' failed",
                       ctx_params.model_path.c_str(),
@@ -766,8 +809,12 @@ int main(int argc, const char* argv[]) {
         if (cli_params.mode == IMG_GEN) {
             sd_img_gen_params_t img_gen_params = gen_params.to_sd_img_gen_params_t();
 
-            num_results = gen_params.batch_count;
-            results.adopt(generate_image(sd_ctx.get(), &img_gen_params), num_results);
+            sd_image_t* generated_images = nullptr;
+            if (!generate_image(sd_ctx.get(), &img_gen_params, &generated_images, &num_results)) {
+                generated_images = nullptr;
+                num_results      = 0;
+            }
+            results.adopt(generated_images, num_results);
         } else if (cli_params.mode == VID_GEN) {
             sd_vid_gen_params_t vid_gen_params = gen_params.to_sd_vid_gen_params_t();
             sd_image_t* generated_video        = nullptr;
@@ -802,12 +849,22 @@ int main(int argc, const char* argv[]) {
                 SDImageOwner current_image(results[i]);
                 results[i] = {0, 0, 0, nullptr};
                 for (int u = 0; u < gen_params.upscale_repeats; ++u) {
-                    SDImageOwner upscaled_image(upscale(upscaler_ctx.get(), current_image.get(), upscale_factor));
-                    if (upscaled_image.get().data == nullptr) {
+                    sd_image_t* upscaled_images = nullptr;
+                    int upscaled_count          = 0;
+                    bool upscale_ok             = upscale(upscaler_ctx.get(),
+                                                          current_image.get(),
+                                                          upscale_factor,
+                                                          &upscaled_images,
+                                                          &upscaled_count);
+                    if (!upscale_ok || upscaled_count <= 0 || upscaled_images[0].data == nullptr) {
+                        free_sd_images(upscaled_images, upscaled_count);
                         LOG_ERROR("upscale failed");
                         break;
                     }
-                    current_image = std::move(upscaled_image);
+                    sd_image_t upscaled_image = upscaled_images[0];
+                    upscaled_images[0]        = {0, 0, 0, nullptr};
+                    free_sd_images(upscaled_images, upscaled_count);
+                    current_image.reset(upscaled_image);
                 }
                 results[i] = current_image.release();  // Set the final upscaled image as the result
             }
@@ -819,6 +876,11 @@ int main(int argc, const char* argv[]) {
         return 1;
     }
 
+    if (!cli_params.imatrix_out.empty()) {
+        LOG_INFO("saving imatrix to '%s'", cli_params.imatrix_out.c_str());
+        save_imatrix(cli_params.imatrix_out.c_str());
+    }
+
     free_sd_audio(generated_audio);
 
     return 0;

examples/common/common.cpp

@@ -468,6 +468,13 @@ ArgOptions SDContextParams::get_options() {
          "parameter backend assignment, e.g. disk, cpu, or diffusion=disk,clip=cpu",
          (int)',',
          &params_backend},
+        {"",
+         "--split-mode",
+         "weight distribution for modules assigned multiple devices (--backend \"diffusion=cuda0&cuda1\"): "
+         "layer (whole transformer blocks per device, default) or row (matmul rows split across devices, CUDA only). "
+         "Accepts a single mode or per-module assignments, e.g. row or diffusion=row,te=layer",
+         (int)',',
+         &split_mode},
         {"",
          "--rpc-servers",
          "comma-separated list of RPC servers to connect to for offloading, in the format host:port, e.g. localhost:50052,192.168.1.3:50052",
@@ -501,6 +508,12 @@ ArgOptions SDContextParams::get_options() {
          "--eager-load",
          "load all params into the params backend at model-load time instead of lazily on first use (defaults to false)",
          true, &eager_load},
+        {"",
+         "--auto-fit",
+         "pick the diffusion/te/vae device placements automatically from the model size and the per-device "
+         "memory budgets (--max-vram; defaults to free memory minus a small margin). Overrides --backend and "
+         "--params-backend; may split modules across GPUs (--split-mode still selects layer or row)",
+         true, &auto_fit},
         {"",
          "--force-sdxl-vae-conv-scale",
          "force use of conv scale on sdxl vae",
@@ -541,18 +554,6 @@ ArgOptions SDContextParams::get_options() {
          "--vae-conv-direct",
          "use ggml_conv2d_direct in the vae model",
          true, &vae_conv_direct},
-        {"",
-         "--circular",
-         "enable circular padding for convolutions",
-         true, &circular},
-        {"",
-         "--circularx",
-         "enable circular RoPE wrapping on x-axis (width) only",
-         true, &circular_x},
-        {"",
-         "--circulary",
-         "enable circular RoPE wrapping on y-axis (height) only",
-         true, &circular_y},
         {"",
          "--chroma-disable-dit-mask",
          "disable dit mask for chroma",
@@ -663,6 +664,18 @@ ArgOptions SDContextParams::get_options() {
          "but it usually offers faster inference speed and, in some cases, lower memory usage. "
          "The at_runtime mode, on the other hand, is exactly the opposite.",
          on_lora_apply_mode_arg},
+        {"",
+         "--list-devices",
+         "list available ggml backend devices (one 'name<TAB>description' per line) and exit; "
+         "the names are the device names accepted by --backend and --params-backend",
+         [](int /*argc*/, const char** /*argv*/, int /*index*/) {
+             size_t device_list_size = sd_list_devices(nullptr, 0);
+             std::vector<char> devices(device_list_size + 1);
+             sd_list_devices(devices.data(), devices.size());
+             fputs(devices.data(), stdout);
+             std::exit(0);
+             return 0;
+         }},
     };
 
     return options;
@@ -710,7 +723,18 @@ bool SDContextParams::resolve(SDMode mode) {
 }
 
 bool SDContextParams::validate(SDMode mode) {
-    if (mode != UPSCALE && mode != METADATA && model_path.length() == 0 && diffusion_model_path.length() == 0) {
+    if (mode == CONVERT) {
+        const bool has_convert_input = model_path.length() != 0 ||
+                                       clip_l_path.length() != 0 ||
+                                       clip_g_path.length() != 0 ||
+                                       t5xxl_path.length() != 0 ||
+                                       diffusion_model_path.length() != 0 ||
+                                       vae_path.length() != 0;
+        if (!has_convert_input) {
+            LOG_ERROR("error: convert mode needs at least one model input path\n");
+            return false;
+        }
+    } else if (mode != UPSCALE && mode != METADATA && model_path.length() == 0 && diffusion_model_path.length() == 0) {
         LOG_ERROR("error: the following arguments are required: model_path/diffusion_model\n");
         return false;
     }
@@ -807,6 +831,8 @@ std::string SDContextParams::to_string() const {
         << "  eager_load: " << (eager_load ? "true" : "false") << ",\n"
         << "  backend: \"" << backend << "\",\n"
         << "  params_backend: \"" << params_backend << "\",\n"
+        << "  split_mode: \"" << split_mode << "\",\n"
+        << "  auto_fit: " << (auto_fit ? "true" : "false") << ",\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"
@@ -815,9 +841,6 @@ std::string SDContextParams::to_string() const {
         << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
         << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
         << "  vae_conv_direct: " << (vae_conv_direct ? "true" : "false") << ",\n"
-        << "  circular: " << (circular ? "true" : "false") << ",\n"
-        << "  circular_x: " << (circular_x ? "true" : "false") << ",\n"
-        << "  circular_y: " << (circular_y ? "true" : "false") << ",\n"
         << "  chroma_use_dit_mask: " << (chroma_use_dit_mask ? "true" : "false") << ",\n"
         << "  qwen_image_zero_cond_t: " << (qwen_image_zero_cond_t ? "true" : "false") << ",\n"
         << "  chroma_use_t5_mask: " << (chroma_use_t5_mask ? "true" : "false") << ",\n"
@@ -874,8 +897,6 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.tae_preview_only                = taesd_preview;
     sd_ctx_params.diffusion_conv_direct           = diffusion_conv_direct;
     sd_ctx_params.vae_conv_direct                 = vae_conv_direct;
-    sd_ctx_params.circular_x                      = circular || circular_x;
-    sd_ctx_params.circular_y                      = circular || circular_y;
     sd_ctx_params.force_sdxl_vae_conv_scale       = force_sdxl_vae_conv_scale;
     sd_ctx_params.chroma_use_dit_mask             = chroma_use_dit_mask;
     sd_ctx_params.chroma_use_t5_mask              = chroma_use_t5_mask;
@@ -887,6 +908,8 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.eager_load                      = eager_load;
     sd_ctx_params.backend                         = effective_backend.c_str();
     sd_ctx_params.params_backend                  = effective_params_backend.c_str();
+    sd_ctx_params.split_mode                      = split_mode.c_str();
+    sd_ctx_params.auto_fit                        = auto_fit;
     sd_ctx_params.rpc_servers                     = rpc_servers.c_str();
     return sd_ctx_params;
 }
@@ -996,6 +1019,10 @@ ArgOptions SDGenerationParams::get_options() {
          "--batch-count",
          "batch count",
          &batch_count},
+        {"",
+         "--qwen-image-layers",
+         "number of Qwen Image Layered layers; latent/output count is layers + 1 (default: 3)",
+         &qwen_image_layers},
         {"",
          "--video-frames",
          "video frames (default: 1)",
@@ -1062,7 +1089,7 @@ ArgOptions SDGenerationParams::get_options() {
          &sample_params.guidance.slg.layer_end},
         {"",
          "--eta",
-         "noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)",
+         "noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde, dpm++2s_a, dpm++2m_sde and dpm++2m_sde_bt)",
          &sample_params.eta},
         {"",
          "--flow-shift",
@@ -1094,7 +1121,7 @@ ArgOptions SDGenerationParams::get_options() {
          &high_noise_sample_params.guidance.slg.layer_end},
         {"",
          "--high-noise-eta",
-         "(high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)",
+         "(high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde, dpm++2s_a, dpm++2m_sde and dpm++2m_sde_bt)",
          &high_noise_sample_params.eta},
         {"",
          "--strength",
@@ -1145,6 +1172,18 @@ ArgOptions SDGenerationParams::get_options() {
          "disable auto resize of ref images",
          false,
          &auto_resize_ref_image},
+        {"",
+         "--circular",
+         "enable circular padding on both axes for tileable output",
+         true, &circular},
+        {"",
+         "--circularx",
+         "enable circular padding on x-axis (width) only",
+         true, &circular_x},
+        {"",
+         "--circulary",
+         "enable circular padding on y-axis (height) only",
+         true, &circular_y},
         {"",
          "--disable-image-metadata",
          "do not embed generation metadata on image files",
@@ -1465,17 +1504,17 @@ 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, dpm++2m_sde, dpm++2m_sde_bt, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
          on_sample_method_arg},
         {"",
          "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, 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, dpm++2m_sde, dpm++2m_sde_bt, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          " default: euler for Flux/SD3/Wan, euler_a otherwise",
          on_high_noise_sample_method_arg},
         {"",
          "--scheduler",
-         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2, logit_normal, flux2, flux], default: model-specific",
+         "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent, ltx2, logit_normal, flux2, flux, beta], alias: normal=discrete, default: model-specific",
          on_scheduler_arg},
         {"",
          "--sigmas",
@@ -1816,6 +1855,7 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("width", width);
     load_if_exists("height", height);
     load_if_exists("batch_count", batch_count);
+    load_if_exists("qwen_image_layers", qwen_image_layers);
     load_if_exists("video_frames", video_frames);
     load_if_exists("fps", fps);
     load_if_exists("upscale_repeats", upscale_repeats);
@@ -2240,6 +2280,11 @@ bool SDGenerationParams::validate(SDMode mode) {
         return false;
     }
 
+    if (qwen_image_layers < 0) {
+        LOG_ERROR("error: qwen_image_layers must be non-negative");
+        return false;
+    }
+
     if (sample_params.sample_steps <= 0) {
         LOG_ERROR("error: the sample_steps must be greater than 0\n");
         return false;
@@ -2406,6 +2451,7 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.strength              = strength;
     params.seed                  = seed;
     params.batch_count           = batch_count;
+    params.qwen_image_layers     = qwen_image_layers;
     params.control_image         = control_image.get();
     params.control_strength      = control_strength;
     params.pm_params             = pm_params;
@@ -2424,6 +2470,8 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.hires.upscale_tile_size   = hires_upscale_tile_size;
     params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
     params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
+    params.circular_x                = circular || circular_x;
+    params.circular_y                = circular || circular_y;
     return params;
 }
 
@@ -2489,6 +2537,8 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     params.hires.upscale_tile_size   = hires_upscale_tile_size;
     params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
     params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
+    params.circular_x                = circular || circular_x;
+    params.circular_y                = circular || circular_y;
     return params;
 }
 
@@ -2531,6 +2581,7 @@ std::string SDGenerationParams::to_string() const {
         << "  width: " << width << ",\n"
         << "  height: " << height << ",\n"
         << "  batch_count: " << batch_count << ",\n"
+        << "  qwen_image_layers: " << qwen_image_layers << ",\n"
         << "  init_image_path: \"" << init_image_path << "\",\n"
         << "  end_image_path: \"" << end_image_path << "\",\n"
         << "  mask_image_path: \"" << mask_image_path << "\",\n"

examples/common/common.h

@@ -151,6 +151,8 @@ struct SDContextParams {
     bool eager_load             = false;
     std::string backend;
     std::string params_backend;
+    std::string split_mode;
+    bool auto_fit = false;
     std::string rpc_servers;
     std::string effective_backend;
     std::string effective_params_backend;
@@ -163,10 +165,6 @@ struct SDContextParams {
     bool diffusion_conv_direct = false;
     bool vae_conv_direct       = false;
 
-    bool circular   = false;
-    bool circular_x = false;
-    bool circular_y = false;
-
     bool chroma_use_dit_mask = true;
     bool chroma_use_t5_mask  = false;
     int chroma_t5_mask_pad   = 1;
@@ -197,6 +195,7 @@ struct SDGenerationParams {
     int width                  = -1;
     int height                 = -1;
     int batch_count            = 1;
+    int qwen_image_layers      = 3;
     int64_t seed               = 42;
     float strength             = 0.75f;
     float control_strength     = 0.9f;
@@ -243,6 +242,10 @@ struct SDGenerationParams {
     int upscale_repeats   = 1;
     int upscale_tile_size = 128;
 
+    bool circular   = false;
+    bool circular_x = false;
+    bool circular_y = false;
+
     bool hires_enabled         = false;
     std::string hires_upscaler = "Latent";
     std::string hires_upscaler_model_path;

examples/server/async_jobs.cpp

@@ -2,6 +2,7 @@
 
 #include "async_jobs.h"
 
+#include <algorithm>
 #include <iomanip>
 #include <sstream>
 
@@ -173,8 +174,13 @@ bool execute_img_gen_job(ServerRuntime& runtime,
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &params);
-        results.adopt(raw_results, params.batch_count);
+        sd_image_t* raw_results = nullptr;
+        int num_results         = 0;
+        if (!generate_image(runtime.sd_ctx, &params, &raw_results, &num_results)) {
+            raw_results = nullptr;
+            num_results = 0;
+        }
+        results.adopt(raw_results, num_results);
     }
 
     const int num_results = results.count();
@@ -190,6 +196,8 @@ bool execute_img_gen_job(ServerRuntime& runtime,
         encoded_format = EncodedImageFormat::WEBP;
     }
 
+    int batch_count      = job.img_gen.gen_params.batch_count;
+    int images_per_batch = batch_count > 0 ? std::max(1, num_results / batch_count) : 1;
     for (int i = 0; i < num_results; ++i) {
         if (results[i].data == nullptr) {
             continue;
@@ -198,7 +206,7 @@ bool execute_img_gen_job(ServerRuntime& runtime,
         const std::string metadata = job.img_gen.gen_params.embed_image_metadata
                                          ? get_image_params(*runtime.ctx_params,
                                                             job.img_gen.gen_params,
-                                                            job.img_gen.gen_params.seed + i)
+                                                            job.img_gen.gen_params.seed + i / images_per_batch)
                                          : "";
         auto image_bytes           = encode_image_to_vector(encoded_format,
                                                             results[i].data,

examples/server/routes_openai.cpp

@@ -229,8 +229,11 @@ static bool execute_sync_img_gen_request(ServerRuntime& runtime,
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
-        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &img_gen_params);
-        num_results             = request.gen_params.batch_count;
+        sd_image_t* raw_results = nullptr;
+        if (!generate_image(runtime.sd_ctx, &img_gen_params, &raw_results, &num_results)) {
+            raw_results = nullptr;
+            num_results = 0;
+        }
         results.adopt(raw_results, num_results);
     }
 
@@ -281,14 +284,16 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
             out["data"]          = json::array();
             out["output_format"] = request.output_format;
 
-            for (int i = 0; i < request.gen_params.batch_count; ++i) {
+            int result_count     = results.count();
+            int images_per_batch = request.gen_params.batch_count > 0 ? std::max(1, result_count / request.gen_params.batch_count) : 1;
+            for (int i = 0; i < result_count; ++i) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
                 std::string params = request.gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params,
                                                             request.gen_params,
-                                                            request.gen_params.seed + i)
+                                                            request.gen_params.seed + i / images_per_batch)
                                          : "";
                 auto image_bytes   = encode_image_to_vector(request.output_format == "jpeg"
                                                                 ? EncodedImageFormat::JPEG
@@ -353,14 +358,16 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
             out["data"]          = json::array();
             out["output_format"] = request.output_format;
 
-            for (int i = 0; i < request.gen_params.batch_count; ++i) {
+            int result_count     = results.count();
+            int images_per_batch = request.gen_params.batch_count > 0 ? std::max(1, result_count / request.gen_params.batch_count) : 1;
+            for (int i = 0; i < result_count; ++i) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
                 std::string params = request.gen_params.embed_image_metadata
                                          ? get_image_params(*runtime->ctx_params,
                                                             request.gen_params,
-                                                            request.gen_params.seed + i)
+                                                            request.gen_params.seed + i / images_per_batch)
                                          : "";
                 auto image_bytes   = encode_image_to_vector(request.output_format == "jpeg" ? EncodedImageFormat::JPEG : EncodedImageFormat::PNG,
                                                           results[i].data,

examples/server/routes_sdapi.cpp

@@ -63,6 +63,10 @@ static enum sample_method_t get_sdapi_sample_method(std::string name) {
         {"ddim", DDIM_TRAILING_SAMPLE_METHOD},
         {"dpm++ 2m", DPMPP2M_SAMPLE_METHOD},
         {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD},
+        {"dpm++ 2m sde", DPMPP2M_SDE_SAMPLE_METHOD},
+        {"k_dpmpp_2m_sde", DPMPP2M_SDE_SAMPLE_METHOD},
+        {"dpm++ 2m sde gpu", DPMPP2M_SDE_BT_SAMPLE_METHOD},
+        {"k_dpmpp_2m_sde_gpu", DPMPP2M_SDE_BT_SAMPLE_METHOD},
         {"res multistep", RES_MULTISTEP_SAMPLE_METHOD},
         {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
         {"res 2s", RES_2S_SAMPLE_METHOD},
@@ -259,6 +263,48 @@ static bool build_sdapi_img_gen_request(const json& j,
     return true;
 }
 
+static nlohmann::json prepare_info_field(const SDContextParams& ctx_params,
+                                         const SDGenerationParams& gen_params,
+                                         bool img2img) {
+    nlohmann::json jsoninfo = nlohmann::json::object();
+    jsoninfo["prompt"]      = gen_params.prompt;
+    if (!gen_params.negative_prompt.empty()) {
+        jsoninfo["negative_prompt"] = gen_params.negative_prompt;
+    }
+    jsoninfo["seed"]         = gen_params.seed;
+    jsoninfo["cfg_scale"]    = gen_params.sample_params.guidance.txt_cfg;
+    jsoninfo["width"]        = gen_params.get_resolved_width();
+    jsoninfo["height"]       = gen_params.get_resolved_height();
+    jsoninfo["steps"]        = gen_params.sample_params.sample_steps;
+    jsoninfo["sampler_name"] = sd_sample_method_name(gen_params.sample_params.sample_method);
+    if (gen_params.clip_skip != -1) {
+        jsoninfo["clip_skip"] = gen_params.clip_skip;
+    }
+    if (gen_params.sample_params.scheduler != scheduler_t::SCHEDULER_COUNT) {
+        jsoninfo["extra_generation_params"]                  = nlohmann::json::object();
+        jsoninfo["extra_generation_params"]["Schedule type"] = sd_scheduler_name(gen_params.sample_params.scheduler);
+    }
+    if (img2img) {
+        jsoninfo["denoising_strength"] = gen_params.strength;
+    }
+    // not clear what should happen if we have both model and diffusion_model
+    if (!ctx_params.diffusion_model_path.empty()) {
+        jsoninfo["sd_model_name"] = sd_basename(ctx_params.diffusion_model_path);
+    } else if (!ctx_params.model_path.empty()) {
+        jsoninfo["sd_model_name"] = sd_basename(ctx_params.model_path);
+    }
+    if (!ctx_params.vae_path.empty()) {
+        jsoninfo["sd_vae_name"] = sd_basename(ctx_params.vae_path);
+    }
+    jsoninfo["version"] = "stable-diffusion.cpp";
+
+    jsoninfo["infotexts"]            = nlohmann::json::array();
+    jsoninfo["all_prompts"]          = nlohmann::json::array();
+    jsoninfo["all_negative_prompts"] = nlohmann::json::array();
+    jsoninfo["all_seeds"]            = nlohmann::json::array();
+    return jsoninfo;
+}
+
 void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
     ServerRuntime* runtime = &rt;
 
@@ -292,8 +338,11 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
             {
                 std::lock_guard<std::mutex> lock(*runtime->sd_ctx_mutex);
-                sd_image_t* raw_results = generate_image(runtime->sd_ctx, &img_gen_params);
-                num_results             = request.gen_params.batch_count;
+                sd_image_t* raw_results = nullptr;
+                if (!generate_image(runtime->sd_ctx, &img_gen_params, &raw_results, &num_results)) {
+                    raw_results = nullptr;
+                    num_results = 0;
+                }
                 results.adopt(raw_results, num_results);
             }
 
@@ -306,24 +355,26 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
             json out;
             out["images"]     = json::array();
             out["parameters"] = j;
-            out["info"]       = "";
+            json jsoninfo     = prepare_info_field(*runtime->ctx_params, request.gen_params, img2img);
 
+            int images_per_batch = request.gen_params.batch_count > 0 ? std::max(1, num_results / request.gen_params.batch_count) : 1;
             for (int i = 0; i < num_results; ++i) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
 
-                std::string params = request.gen_params.embed_image_metadata
-                                         ? get_image_params(*runtime->ctx_params,
-                                                            request.gen_params,
-                                                            request.gen_params.seed + i)
-                                         : "";
-                auto image_bytes   = encode_image_to_vector(EncodedImageFormat::PNG,
-                                                            results[i].data,
-                                                            results[i].width,
-                                                            results[i].height,
-                                                            results[i].channel,
-                                                            params);
+                bool embed_meta = request.gen_params.embed_image_metadata;
+
+                std::string params = get_image_params(*runtime->ctx_params,
+                                                      request.gen_params,
+                                                      request.gen_params.seed + i / images_per_batch);
+
+                auto image_bytes = encode_image_to_vector(EncodedImageFormat::PNG,
+                                                          results[i].data,
+                                                          results[i].width,
+                                                          results[i].height,
+                                                          results[i].channel,
+                                                          embed_meta ? params : "");
 
                 if (image_bytes.empty()) {
                     LOG_ERROR("write image to mem failed");
@@ -331,8 +382,16 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 }
 
                 out["images"].push_back(base64_encode(image_bytes));
+
+                jsoninfo["infotexts"][i]            = params;
+                jsoninfo["all_seeds"][i]            = request.gen_params.seed + i;
+                jsoninfo["all_prompts"][i]          = request.gen_params.prompt;
+                jsoninfo["all_negative_prompts"][i] = request.gen_params.negative_prompt;
             }
 
+            // not a mistake: it is supposed to be a string in json format
+            out["info"] = jsoninfo.dump();
+
             res.set_content(out.dump(), "application/json");
             res.status = 200;
 
@@ -438,6 +497,9 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         scheduler_names.push_back("default");
         for (int i = 0; i < SCHEDULER_COUNT; i++) {
             scheduler_names.push_back(sd_scheduler_name((scheduler_t)i));
+            if (i == DISCRETE_SCHEDULER) {
+                scheduler_names.push_back("normal");
+            }
         }
         json r = json::array();
         for (auto name : scheduler_names) {

examples/server/routes_sdcpp.cpp

@@ -126,6 +126,7 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
         {"strength", defaults.strength},
         {"seed", defaults.seed},
         {"batch_count", defaults.batch_count},
+        {"qwen_image_layers", defaults.qwen_image_layers},
         {"auto_resize_ref_image", defaults.auto_resize_ref_image},
         {"increase_ref_index", defaults.increase_ref_index},
         {"control_strength", defaults.control_strength},
@@ -219,6 +220,9 @@ static json make_capabilities_json(ServerRuntime& runtime) {
 
     for (int i = 0; i < SCHEDULER_COUNT; ++i) {
         schedulers.push_back(sd_scheduler_name((scheduler_t)i));
+        if (i == DISCRETE_SCHEDULER) {
+            schedulers.push_back("normal");
+        }
     }
 
     {

include/stable-diffusion.h

@@ -54,6 +54,8 @@ enum sample_method_t {
     EULER_CFG_PP_SAMPLE_METHOD,
     EULER_A_CFG_PP_SAMPLE_METHOD,
     EULER_GE_SAMPLE_METHOD,
+    DPMPP2M_SDE_SAMPLE_METHOD,
+    DPMPP2M_SDE_BT_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -73,6 +75,7 @@ enum scheduler_t {
     LOGIT_NORMAL_SCHEDULER,
     FLUX2_SCHEDULER,
     FLUX_SCHEDULER,
+    BETA_SCHEDULER,
     SCHEDULER_COUNT
 };
 
@@ -83,6 +86,7 @@ enum prediction_t {
     FLOW_PRED,
     FLUX_FLOW_PRED,
     SEFI_FLOW_PRED,
+    MINIT2I_FLOW_PRED,
     PREDICTION_COUNT
 };
 
@@ -212,8 +216,6 @@ typedef struct {
     bool tae_preview_only;
     bool diffusion_conv_direct;
     bool vae_conv_direct;
-    bool circular_x;
-    bool circular_y;
     bool force_sdxl_vae_conv_scale;
     bool chroma_use_dit_mask;
     bool chroma_use_t5_mask;
@@ -225,6 +227,8 @@ typedef struct {
     bool eager_load;  // Load all params into the params backend at model-load time instead of lazily on first use
     const char* backend;
     const char* params_backend;
+    const char* split_mode;  // weight distribution for multi-device modules: layer (default) or row, or per-module assignments e.g. "diffusion=row"
+    bool auto_fit;
     const char* rpc_servers;
 } sd_ctx_params_t;
 
@@ -378,6 +382,9 @@ typedef struct {
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
     sd_hires_params_t hires;
+    int qwen_image_layers;
+    bool circular_x;
+    bool circular_y;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -403,17 +410,22 @@ typedef struct {
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
     sd_hires_params_t hires;
+    bool circular_x;
+    bool circular_y;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
+struct ggml_tensor;
 
 typedef void (*sd_log_cb_t)(enum sd_log_level_t level, const char* text, void* data);
 typedef void (*sd_progress_cb_t)(int step, int steps, float time, void* data);
 typedef void (*sd_preview_cb_t)(int step, int frame_count, sd_image_t* frames, bool is_noisy, void* data);
+typedef bool (*sd_graph_eval_callback_t)(struct ggml_tensor* t, bool ask, void* user_data);
 
 SD_API void sd_set_log_callback(sd_log_cb_t sd_log_cb, void* data);
 SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
 SD_API void sd_set_preview_callback(sd_preview_cb_t cb, enum preview_t mode, int interval, bool denoised, bool noisy, void* data);
+SD_API void sd_set_backend_eval_callback(sd_graph_eval_callback_t cb, void* data);
 SD_API int32_t sd_get_num_physical_cores();
 SD_API const char* sd_get_system_info();
 SD_API bool sd_ctx_supports_image_generation(const sd_ctx_t* sd_ctx);
@@ -454,7 +466,10 @@ SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx, enum sa
 
 SD_API void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params);
 SD_API char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params);
-SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params);
+SD_API bool generate_image(sd_ctx_t* sd_ctx,
+                           const sd_img_gen_params_t* sd_img_gen_params,
+                           sd_image_t** images_out,
+                           int* num_images_out);
 
 enum sd_cancel_mode_t {
     // Stop the current generation as soon as possible.
@@ -484,9 +499,11 @@ SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
                                         const char* params_backend);
 SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
 
-SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,
-                          sd_image_t input_image,
-                          uint32_t upscale_factor);
+SD_API bool upscale(upscaler_ctx_t* upscaler_ctx,
+                    sd_image_t input_image,
+                    uint32_t upscale_factor,
+                    sd_image_t** images_out,
+                    int* num_images_out);
 
 SD_API int get_upscale_factor(upscaler_ctx_t* upscaler_ctx);
 
@@ -497,6 +514,18 @@ SD_API bool convert(const char* input_path,
                     const char* tensor_type_rules,
                     bool convert_name);
 
+SD_API bool convert_with_components(const char* model_path,
+                                    const char* clip_l_path,
+                                    const char* clip_g_path,
+                                    const char* t5xxl_path,
+                                    const char* diffusion_model_path,
+                                    const char* vae_path,
+                                    const char* output_path,
+                                    enum sd_type_t output_type,
+                                    const char* tensor_type_rules,
+                                    bool convert_name,
+                                    int n_threads);
+
 SD_API bool preprocess_canny(sd_image_t image,
                              float high_threshold,
                              float low_threshold,
@@ -504,9 +533,20 @@ SD_API bool preprocess_canny(sd_image_t image,
                              float strong,
                              bool inverse);
 
+SD_API bool load_imatrix(const char* imatrix_path);
+SD_API void save_imatrix(const char* imatrix_path);
+SD_API void enable_imatrix_collection(void);
+SD_API void disable_imatrix_collection(void);
+
 SD_API const char* sd_commit(void);
 SD_API const char* sd_version(void);
 
+// List available ggml backend devices, one `name<TAB>description` per line.
+// The names are the device names accepted by the --backend / --params-backend
+// assignment specs. Returns the number of bytes required, excluding the null
+// terminator. Passing nullptr or buffer_size 0 only queries the required size.
+SD_API size_t sd_list_devices(char* buffer, size_t buffer_size);
+
 // for C API, caller needs to call free_sd_images to free the memory after use
 // This helps avoid CRT problems on Windows when memory is allocated in the library but freed in the caller, which may use a different CRT.
 SD_API void free_sd_images(sd_image_t* result_images, int num_images);

scripts/convert_qwen3_vl.py

@@ -9,7 +9,7 @@ the text side; the vision side is converted by sd.cpp's own
 Operates on raw safetensors bytes so any dtype (BF16/F16/F32) is preserved.
 
 Usage:
-    python3 script/convert_qwen3_vl.py <hf_qwen3_vl_dir_or_safetensors> <output.safetensors>
+    python3 scripts/convert_qwen3_vl.py <hf_qwen3_vl_dir_or_safetensors> <output.safetensors>
 """
 
 import argparse

scripts/convert_sefi.py

@@ -5,7 +5,7 @@ Operates on raw safetensors bytes so any dtype (BF16, F32, ...) is preserved exa
 No numpy or torch dependency required.
 
 Usage:
-    python3 script/convert_sefi.py <sefi_diffusers_dir> <output.safetensors>
+    python3 scripts/convert_sefi.py <sefi_diffusers_dir> <output.safetensors>
 """
 
 import argparse

scripts/format-code.ps1

@@ -31,24 +31,30 @@ $patterns = @(
     "examples/common/*.cpp"
 )
 
-$root = (Get-Location).Path
+Push-Location (Join-Path $PSScriptRoot "..")
 
-foreach ($pattern in $patterns) {
-    $files = Get-ChildItem -Path $pattern -File -ErrorAction SilentlyContinue | Sort-Object FullName
+try {
+    $root = (Get-Location).Path
 
-    foreach ($file in $files) {
-        $relativePath = $file.FullName.Substring($root.Length).TrimStart('\', '/') -replace '\\', '/'
+    foreach ($pattern in $patterns) {
+        $files = Get-ChildItem -Path $pattern -File -ErrorAction SilentlyContinue | Sort-Object FullName
 
-        if ($relativePath -like "vocab*") {
-            continue
+        foreach ($file in $files) {
+            $relativePath = $file.FullName.Substring($root.Length).TrimStart('\', '/') -replace '\\', '/'
+
+            if ($relativePath -like "vocab*") {
+                continue
+            }
+
+            Write-Host "formatting '$relativePath'"
+
+            # if ($relativePath -ne "stable-diffusion.h") {
+            #     clang-tidy -fix -p build_linux/ "$relativePath"
+            # }
+
+            & clang-format -style=file -i $relativePath
         }
-
-        Write-Host "formatting '$relativePath'"
-
-        # if ($relativePath -ne "stable-diffusion.h") {
-        #     clang-tidy -fix -p build_linux/ "$relativePath"
-        # }
-
-        & clang-format -style=file -i $relativePath
     }
+} finally {
+    Pop-Location
 }

scripts/format-code.sh

@@ -1,3 +1,10 @@
+#!/usr/bin/env bash
+
+set -e
+
+repo_root="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$repo_root"
+
 for f in src/*.cpp src/*.h src/*.hpp \
          src/conditioning/*.cpp src/conditioning/*.h src/conditioning/*.hpp \
          src/core/*.cpp src/core/*.h src/core/*.hpp \

scripts/pulid_extract_id.py

@@ -13,7 +13,7 @@ packaging quirks):
   - numpy, Pillow
 
 Usage:
-  python script/pulid_extract_id.py \\
+  python scripts/pulid_extract_id.py \\
     --portrait /path/to/source-photo.jpg \\
     --pulid-weights /path/to/pulid_flux_v0.9.1.safetensors \\
     --out /path/to/source.pulidembd

scripts/remove_utf8_bom.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+"""Remove UTF-8 BOMs from files under a directory.
+
+By default this scans the current working directory recursively and skips
+repository areas that should not be touched by ordinary maintenance scripts.
+Only files whose first three bytes are the UTF-8 BOM are rewritten.
+"""
+
+import argparse
+import os
+import shutil
+import sys
+import tempfile
+from pathlib import Path
+
+
+UTF8_BOM = b"\xef\xbb\xbf"
+
+DEFAULT_EXCLUDED_DIR_NAMES = {
+    ".git",
+    ".hg",
+    ".svn",
+    ".mypy_cache",
+    ".pytest_cache",
+    "__pycache__",
+    "test",
+}
+
+DEFAULT_EXCLUDED_DIR_PREFIXES = {
+    "build",
+}
+
+DEFAULT_EXCLUDED_REL_DIRS = {
+    "examples/server/frontend",
+    "ggml",
+    "models",
+    "src/vocab",
+    "thirdparty",
+}
+
+
+def rel_posix(path: Path, root: Path) -> str:
+    try:
+        return path.relative_to(root).as_posix()
+    except ValueError:
+        return path.as_posix()
+
+
+def should_skip_dir(
+    path: Path,
+    root: Path,
+    excluded_rel_dirs: set[str],
+    excluded_names: set[str],
+    excluded_prefixes: set[str],
+) -> bool:
+    rel = rel_posix(path, root)
+    return (
+        path.name in excluded_names
+        or rel in excluded_rel_dirs
+        or any(path.name.startswith(prefix) for prefix in excluded_prefixes)
+    )
+
+
+def iter_files(
+    root: Path,
+    recursive: bool,
+    excluded_rel_dirs: set[str],
+    excluded_names: set[str],
+    excluded_prefixes: set[str],
+    follow_symlinks: bool,
+):
+    if recursive:
+        for dirpath, dirnames, filenames in os.walk(root, followlinks=follow_symlinks):
+            current_dir = Path(dirpath)
+            dirnames[:] = [
+                name
+                for name in dirnames
+                if not should_skip_dir(
+                    current_dir / name,
+                    root,
+                    excluded_rel_dirs,
+                    excluded_names,
+                    excluded_prefixes,
+                )
+            ]
+            for filename in filenames:
+                path = current_dir / filename
+                if path.is_symlink() and not follow_symlinks:
+                    continue
+                yield path
+    else:
+        for path in root.iterdir():
+            if path.is_file() and (follow_symlinks or not path.is_symlink()):
+                yield path
+
+
+def has_utf8_bom(path: Path) -> bool:
+    with path.open("rb") as f:
+        return f.read(len(UTF8_BOM)) == UTF8_BOM
+
+
+def strip_utf8_bom(path: Path) -> None:
+    tmp_path = None
+    try:
+        with path.open("rb") as src:
+            if src.read(len(UTF8_BOM)) != UTF8_BOM:
+                return
+
+            fd, tmp_name = tempfile.mkstemp(
+                prefix=f".{path.name}.",
+                suffix=".tmp",
+                dir=str(path.parent),
+            )
+            tmp_path = Path(tmp_name)
+            with os.fdopen(fd, "wb") as dst:
+                shutil.copyfileobj(src, dst, length=1024 * 1024)
+
+        shutil.copystat(path, tmp_path, follow_symlinks=False)
+        os.replace(tmp_path, path)
+        tmp_path = None
+    finally:
+        if tmp_path is not None:
+            try:
+                tmp_path.unlink()
+            except FileNotFoundError:
+                pass
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Scan files and convert UTF-8 BOM files to UTF-8 without BOM.",
+    )
+    parser.add_argument(
+        "root",
+        nargs="?",
+        default=".",
+        help="Directory to scan. Defaults to the current directory.",
+    )
+    parser.add_argument(
+        "-n",
+        "--dry-run",
+        action="store_true",
+        help="Only list files that would be converted.",
+    )
+    parser.add_argument(
+        "--no-recursive",
+        action="store_true",
+        help="Only scan files directly under root.",
+    )
+    parser.add_argument(
+        "--include-repo-excluded",
+        action="store_true",
+        help="Do not skip default repository excluded directories.",
+    )
+    parser.add_argument(
+        "--exclude-dir",
+        action="append",
+        default=[],
+        metavar="DIR",
+        help="Additional directory name or root-relative path to skip. Can be used multiple times.",
+    )
+    parser.add_argument(
+        "--follow-symlinks",
+        action="store_true",
+        help="Follow symlinked directories and files.",
+    )
+    parser.add_argument(
+        "-q",
+        "--quiet",
+        action="store_true",
+        help="Only print the final summary.",
+    )
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    root = Path(args.root).resolve()
+
+    if not root.is_dir():
+        print(f"error: not a directory: {root}", file=sys.stderr)
+        return 2
+
+    excluded_names = set()
+    excluded_rel_dirs = set()
+    excluded_prefixes = set()
+    if not args.include_repo_excluded:
+        excluded_names.update(DEFAULT_EXCLUDED_DIR_NAMES)
+        excluded_rel_dirs.update(DEFAULT_EXCLUDED_REL_DIRS)
+        excluded_prefixes.update(DEFAULT_EXCLUDED_DIR_PREFIXES)
+
+    for item in args.exclude_dir:
+        normalized = Path(item).as_posix().strip("/")
+        if "/" in normalized:
+            excluded_rel_dirs.add(normalized)
+        else:
+            excluded_names.add(normalized)
+
+    scanned = 0
+    converted = 0
+    errors = 0
+
+    for path in iter_files(
+        root=root,
+        recursive=not args.no_recursive,
+        excluded_rel_dirs=excluded_rel_dirs,
+        excluded_names=excluded_names,
+        excluded_prefixes=excluded_prefixes,
+        follow_symlinks=args.follow_symlinks,
+    ):
+        scanned += 1
+        try:
+            if not has_utf8_bom(path):
+                continue
+            converted += 1
+            rel = rel_posix(path, root)
+            if args.dry_run:
+                if not args.quiet:
+                    print(f"would convert: {rel}")
+            else:
+                strip_utf8_bom(path)
+                if not args.quiet:
+                    print(f"converted: {rel}")
+        except OSError as exc:
+            errors += 1
+            print(f"error: {rel_posix(path, root)}: {exc}", file=sys.stderr)
+
+    action = "would convert" if args.dry_run else "converted"
+    print(f"scanned {scanned} file(s), {action} {converted}, errors {errors}")
+    return 1 if errors else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/conditioning/conditioner.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_CONDITIONING_CONDITIONER_HPP__
+#ifndef __SD_CONDITIONING_CONDITIONER_HPP__
 #define __SD_CONDITIONING_CONDITIONER_HPP__
 
 #include <cmath>
@@ -116,6 +116,8 @@ public:
     virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)           = 0;
     virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
     virtual void set_stream_layers_enabled(bool enabled) {}
+    virtual void set_runtime_backends(const std::vector<ggml_backend_t>& backends) {}
+    virtual void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {}
     virtual void set_flash_attention_enabled(bool enabled) = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual void runner_done() {}
@@ -635,6 +637,18 @@ struct SD3CLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -994,6 +1008,18 @@ struct FluxCLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1226,6 +1252,18 @@ struct T5CLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (t5) {
             t5->set_flash_attention_enabled(enabled);
@@ -1378,6 +1416,113 @@ struct T5CLIPEmbedder : public Conditioner {
     }
 };
 
+struct MiniT2IConditioner : public Conditioner {
+    T5UniGramTokenizer tokenizer;
+    std::shared_ptr<T5Runner> t5;
+    size_t prompt_length = 256;
+
+    MiniT2IConditioner(ggml_backend_t backend,
+                       const String2TensorStorage& tensor_storage_map      = {},
+                       std::shared_ptr<RunnerWeightManager> weight_manager = nullptr) {
+        bool use_t5 = false;
+        for (const auto& pair : tensor_storage_map) {
+            if (pair.first.find("text_encoders.t5xxl") != std::string::npos) {
+                use_t5 = true;
+                break;
+            }
+        }
+        if (!use_t5) {
+            LOG_WARN("IMPORTANT NOTICE: No MiniT2I T5 text encoder provided, cannot process prompts!");
+            return;
+        }
+        t5 = std::make_shared<T5Runner>(backend, tensor_storage_map, "text_encoders.t5xxl.transformer", false, weight_manager);
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
+    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_stream_layers_enabled(bool enabled) override {
+        if (t5) {
+            t5->set_stream_layers_enabled(enabled);
+        }
+    }
+
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        if (t5) {
+            t5->set_runtime_backends(backends);
+        }
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (t5) {
+            t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+        }
+    }
+
+    void set_flash_attention_enabled(bool enabled) override {
+        if (t5) {
+            t5->set_flash_attention_enabled(enabled);
+        }
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        if (t5) {
+            t5->set_weight_adapter(adapter);
+        }
+    }
+
+    void runner_done() override {
+        if (t5) {
+            t5->runner_done();
+        }
+    }
+
+    SDCondition get_learned_condition(int n_threads,
+                                      const ConditionerParams& conditioner_params) override {
+        SDCondition result;
+        if (!t5) {
+            result.c_crossattn = sd::Tensor<float>::zeros({1024, static_cast<int64_t>(prompt_length)});
+            result.c_vector    = sd::Tensor<float>::zeros({static_cast<int64_t>(prompt_length)});
+            return result;
+        }
+
+        std::vector<int> tokens = tokenizer.encode(conditioner_params.text);
+        if (tokens.size() > prompt_length) {
+            tokens.resize(prompt_length);
+        }
+        std::vector<float> mask(tokens.size(), 1.0f);
+        while (tokens.size() < prompt_length) {
+            tokens.push_back(tokenizer.PAD_TOKEN_ID);
+            mask.push_back(0.0f);
+        }
+
+        sd::Tensor<int32_t> input_ids({static_cast<int64_t>(tokens.size())}, tokens);
+        std::vector<float> t5_mask(mask.size(), 0.0f);
+        for (size_t i = 0; i < mask.size(); ++i) {
+            t5_mask[i] = mask[i] > 0.0f ? 0.0f : -HUGE_VALF;
+        }
+        sd::Tensor<float> hidden_states = t5->compute(n_threads,
+                                                      input_ids,
+                                                      sd::Tensor<float>::from_vector(t5_mask),
+                                                      false,
+                                                      true,
+                                                      true);
+        GGML_ASSERT(!hidden_states.empty());
+        result.c_crossattn = std::move(hidden_states);
+        result.c_vector    = sd::Tensor<float>::from_vector(mask);
+        return result;
+    }
+};
+
 struct AnimaConditioner : public Conditioner {
     std::shared_ptr<BPETokenizer> qwen_tokenizer;
     T5UniGramTokenizer t5_tokenizer;
@@ -1407,6 +1552,14 @@ struct AnimaConditioner : public Conditioner {
         llm->set_stream_layers_enabled(enabled);
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        llm->set_runtime_backends(backends);
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -1552,6 +1705,14 @@ struct LLMEmbedder : public Conditioner {
         llm->set_stream_layers_enabled(enabled);
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        llm->set_runtime_backends(backends);
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -2221,6 +2382,14 @@ struct LTXAVEmbedder : public Conditioner {
         projector->set_max_graph_vram_bytes(max_vram_bytes);
     }
 
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
+        llm->set_runtime_backends(backends);
+    }
+
+    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         llm->set_weight_adapter(adapter);
         projector->set_weight_adapter(adapter);

src/convert.cpp

@@ -1,14 +1,33 @@
+#include <algorithm>
+#include <condition_variable>
+#include <cstdint>
 #include <cstring>
+#include <exception>
+#include <fstream>
+#include <memory>
 #include <mutex>
 #include <regex>
+#include <string>
+#include <thread>
 #include <vector>
 
+#include "core/util.h"
 #include "model_io/gguf_io.h"
 #include "model_io/safetensors_io.h"
+#include "model_io/streaming_writer.h"
 #include "model_loader.h"
-#include "util.h"
 
-#include "ggml_extend_backend.h"
+struct TensorExportInfo {
+    TensorStorage storage;
+    ggml_type type;
+};
+
+struct TensorExportJob {
+    TensorExportInfo info;
+    std::vector<uint8_t> data;
+    std::string error;
+    bool success = false;
+};
 
 static ggml_type get_export_tensor_type(ModelLoader& model_loader,
                                         const TensorStorage& tensor_storage,
@@ -33,106 +52,355 @@ static ggml_type get_export_tensor_type(ModelLoader& model_loader,
     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);
+static bool collect_tensors_for_export(ModelLoader& model_loader,
+                                       ggml_type type,
+                                       const TensorTypeRules& tensor_type_rules,
+                                       std::vector<TensorExportInfo>& tensors) {
+    tensors.clear();
+    tensors.reserve(model_loader.get_tensor_storage_map().size());
+    for (const auto& kv : model_loader.get_tensor_storage_map()) {
+        const TensorStorage& tensor_storage = kv.second;
+        TensorExportInfo info;
+        info.storage = tensor_storage;
+        info.type    = get_export_tensor_type(model_loader, tensor_storage, type, tensor_type_rules);
+        tensors.push_back(std::move(info));
+    }
+    LOG_INFO("collected %zu tensors for export", tensors.size());
+    return true;
+}
 
-        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");
+static size_t export_tensor_nbytes(const TensorExportInfo& info) {
+    TensorStorage output_storage = info.storage;
+    output_storage.type          = info.type;
+    return static_cast<size_t>(output_storage.nbytes());
+}
+
+static TensorWritePlan tensor_write_plan_from_export_info(const TensorExportInfo& info) {
+    TensorWritePlan plan;
+    plan.name   = info.storage.name;
+    plan.type   = info.type;
+    plan.n_dims = info.storage.n_dims;
+    for (int i = 0; i < SD_MAX_DIMS; i++) {
+        plan.ne[i] = info.storage.ne[i];
+    }
+    return plan;
+}
+
+static std::vector<TensorWritePlan> tensor_write_plans_from_export_infos(const std::vector<TensorExportInfo>& tensors) {
+    std::vector<TensorWritePlan> plans;
+    plans.reserve(tensors.size());
+    for (const TensorExportInfo& info : tensors) {
+        plans.push_back(tensor_write_plan_from_export_info(info));
+    }
+    return plans;
+}
+
+static bool preallocate_output_file(const std::string& output_path, uint64_t file_size, std::string* error) {
+    if (file_size == 0) {
+        return true;
+    }
+
+    std::fstream file(output_path, std::ios::binary | std::ios::in | std::ios::out);
+    if (!file.is_open()) {
+        if (error != nullptr) {
+            *error = "failed to open output file '" + output_path + "' for preallocation";
+        }
+        return false;
+    }
+
+    // This portable fallback sets the final file size. A platform-specific
+    // posix_fallocate/ftruncate path can replace it later.
+    file.seekp(static_cast<std::streamoff>(file_size - 1), std::ios::beg);
+    file.put('\0');
+    file.flush();
+    if (!file) {
+        if (error != nullptr) {
+            *error = "failed to preallocate output file '" + output_path + "'";
+        }
+        return false;
+    }
+    return true;
+}
+
+static bool load_tensor_for_export(ModelLoader& model_loader, TensorExportJob& job) {
+    size_t mem_size = 1 * 1024 * 1024;
+    mem_size += ggml_tensor_overhead();
+    TensorStorage output_storage = job.info.storage;
+    output_storage.type          = job.info.type;
+    mem_size += static_cast<size_t>(output_storage.nbytes());
+
+    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
+    if (ggml_ctx == nullptr) {
+        job.error = "ggml_init failed for tensor '" + job.info.storage.name + "'";
+        return false;
+    }
+
+    ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, job.info.type, job.info.storage.n_dims, job.info.storage.ne);
+    if (tensor == nullptr) {
+        ggml_free(ggml_ctx);
+        job.error = "ggml_new_tensor failed for tensor '" + job.info.storage.name + "'";
+        return false;
+    }
+    ggml_set_name(tensor, job.info.storage.name.c_str());
+
+    const size_t tensor_nbytes = ggml_nbytes(tensor);
+    if (tensor_nbytes > 0 && !model_loader.load_tensor(job.info.storage, tensor)) {
+        ggml_free(ggml_ctx);
+        job.error = "failed to load tensor '" + job.info.storage.name + "'";
+        return false;
+    }
+
+    job.data.resize(tensor_nbytes);
+    if (tensor_nbytes > 0) {
+        memcpy(job.data.data(), tensor->data, tensor_nbytes);
+    }
+    ggml_free(ggml_ctx);
+    return true;
+}
+
+static bool stream_tensor_data(ModelLoader& model_loader,
+                               const std::string& output_path,
+                               const std::vector<TensorExportInfo>& tensors,
+                               const StreamingModelWriter& writer,
+                               int n_threads,
+                               std::string* error) {
+    n_threads = n_threads > 0 ? n_threads : sd_get_num_physical_cores();
+    n_threads = std::max(1, n_threads);
+    LOG_INFO("streaming convert with %d threads", n_threads);
+
+    int64_t start_time       = ggml_time_ms();
+    uint64_t bytes_written   = 0;
+    size_t tensors_written   = 0;
+    size_t next_tensor_index = 0;
+    bool failed              = false;
+    std::string failure;
+
+    const size_t memory_budget = 1024ull * 1024ull * 1024ull;
+    size_t reserved_bytes      = 0;
+
+    std::mutex work_mutex;
+    std::mutex progress_mutex;
+    std::condition_variable memory_cv;
+    std::vector<std::thread> workers;
+    workers.reserve(n_threads);
+
+    auto reserve_memory = [&](size_t bytes) -> bool {
+        std::unique_lock<std::mutex> lock(work_mutex);
+        memory_cv.wait(lock, [&]() {
+            return failed || reserved_bytes == 0 || reserved_bytes + bytes <= memory_budget;
+        });
+        if (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));
-
+        reserved_bytes += bytes;
         return true;
     };
 
-    bool success = model_loader.load_tensors(on_new_tensor_cb);
-    LOG_INFO("load tensors done");
+    auto release_memory = [&](size_t bytes) {
+        {
+            std::lock_guard<std::mutex> lock(work_mutex);
+            reserved_bytes -= std::min(reserved_bytes, bytes);
+        }
+        memory_cv.notify_all();
+    };
+
+    auto fail = [&](const std::string& message) {
+        {
+            std::lock_guard<std::mutex> lock(work_mutex);
+            if (!failed) {
+                failed  = true;
+                failure = message;
+            }
+        }
+        memory_cv.notify_all();
+    };
+
+    for (int worker = 0; worker < n_threads; worker++) {
+        workers.emplace_back([&]() {
+            std::fstream output_file(output_path, std::ios::binary | std::ios::in | std::ios::out);
+            if (!output_file.is_open()) {
+                fail("failed to open output file '" + output_path + "' for tensor writing");
+                return;
+            }
+
+            while (true) {
+                size_t tensor_index = 0;
+                {
+                    std::lock_guard<std::mutex> lock(work_mutex);
+                    if (failed || next_tensor_index >= tensors.size()) {
+                        return;
+                    }
+                    tensor_index = next_tensor_index++;
+                }
+
+                const size_t tensor_bytes = export_tensor_nbytes(tensors[tensor_index]);
+                if (!reserve_memory(tensor_bytes)) {
+                    return;
+                }
+
+                TensorExportJob job;
+                job.info = tensors[tensor_index];
+                try {
+                    job.success = load_tensor_for_export(model_loader, job);
+                } catch (const std::exception& e) {
+                    job.error   = e.what();
+                    job.success = false;
+                }
+
+                if (!job.success) {
+                    release_memory(tensor_bytes);
+                    fail(job.error.empty() ? "streaming conversion failed" : job.error);
+                    return;
+                }
+
+                std::string write_error;
+                if (!writer.write_tensor(output_file,
+                                         tensor_index,
+                                         job.data.empty() ? nullptr : job.data.data(),
+                                         job.data.size(),
+                                         &write_error)) {
+                    release_memory(tensor_bytes);
+                    fail(write_error.empty() ? "streaming conversion write failed" : write_error);
+                    return;
+                }
+
+                {
+                    std::lock_guard<std::mutex> lock(progress_mutex);
+                    bytes_written += job.data.size();
+                    tensors_written++;
+                    float elapsed_seconds = (ggml_time_ms() - start_time) / 1000.0f;
+                    pretty_bytes_progress(static_cast<int>(tensors_written),
+                                          static_cast<int>(tensors.size()),
+                                          bytes_written,
+                                          elapsed_seconds);
+                }
+                release_memory(tensor_bytes);
+            }
+        });
+    }
+
+    for (auto& worker : workers) {
+        worker.join();
+    }
+    printf("\n");
+    if (failed) {
+        if (error != nullptr) {
+            *error = failure;
+        }
+        return false;
+    }
+    LOG_INFO("streaming conversion completed, taking %.2fs", (ggml_time_ms() - start_time) / 1000.f);
+    return true;
+}
+
+static bool write_model_file_streaming(ModelLoader& model_loader,
+                                       const std::string& output_path,
+                                       const std::vector<TensorExportInfo>& tensors,
+                                       StreamingModelWriter& writer,
+                                       int n_threads,
+                                       std::string* error) {
+    std::vector<TensorWritePlan> plans = tensor_write_plans_from_export_infos(tensors);
+    if (!writer.write_metadata(output_path, plans, error)) {
+        return false;
+    }
+    if (!preallocate_output_file(output_path, writer.file_size(), error)) {
+        return false;
+    }
+    model_loader.process_model_files(false, false);
+    return stream_tensor_data(model_loader, output_path, tensors, writer, n_threads, error);
+}
+
+static bool init_convert_path(ModelLoader& model_loader, const char* path, const char* prefix, bool& loaded_any) {
+    if (path == nullptr || strlen(path) == 0) {
+        return true;
+    }
+    if (!model_loader.init_from_file(path, prefix)) {
+        LOG_ERROR("init model loader from file failed: '%s'", path);
+        return false;
+    }
+    loaded_any = true;
+    return true;
+}
+
+static bool export_loaded_model(ModelLoader& model_loader,
+                                const char* output_path,
+                                sd_type_t output_type,
+                                const char* tensor_type_rules,
+                                int n_threads) {
+    ggml_type type             = sd_type_to_ggml_type(output_type);
+    bool output_is_safetensors = ends_with(output_path, ".safetensors");
+    TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);
+
+    std::vector<TensorExportInfo> tensors;
+    bool success = collect_tensors_for_export(model_loader, type, type_rules, tensors);
+    std::string error;
+    if (success) {
+        std::unique_ptr<StreamingModelWriter> writer;
+        if (output_is_safetensors) {
+            writer = std::make_unique<SafetensorsStreamingWriter>();
+        } else {
+            writer = std::make_unique<GGUFStreamingWriter>();
+        }
+        success = write_model_file_streaming(model_loader, output_path, tensors, *writer, n_threads, &error);
+    }
+
+    if (!success && !error.empty()) {
+        LOG_ERROR("%s", error.c_str());
+    }
+
     return success;
 }
 
+bool convert_with_components(const char* model_path,
+                             const char* clip_l_path,
+                             const char* clip_g_path,
+                             const char* t5xxl_path,
+                             const char* diffusion_model_path,
+                             const char* vae_path,
+                             const char* output_path,
+                             sd_type_t output_type,
+                             const char* tensor_type_rules,
+                             bool convert_name,
+                             int n_threads) {
+    ModelLoader model_loader;
+    bool loaded_any = false;
+
+    if (!init_convert_path(model_loader, model_path, "", loaded_any) ||
+        !init_convert_path(model_loader, clip_l_path, "text_encoders.clip_l.transformer.", loaded_any) ||
+        !init_convert_path(model_loader, clip_g_path, "text_encoders.clip_g.transformer.", loaded_any) ||
+        !init_convert_path(model_loader, t5xxl_path, "text_encoders.t5xxl.transformer.", loaded_any) ||
+        !init_convert_path(model_loader, diffusion_model_path, "model.diffusion_model.", loaded_any) ||
+        !init_convert_path(model_loader, vae_path, "vae.", loaded_any)) {
+        return false;
+    }
+
+    if (!loaded_any) {
+        LOG_ERROR("no input model path provided for convert");
+        return false;
+    }
+
+    if (convert_name) {
+        model_loader.convert_tensors_name();
+    }
+
+    return export_loaded_model(model_loader, output_path, output_type, tensor_type_rules, n_threads);
+}
+
 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             = sd_type_to_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    = sd_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;
+    return convert_with_components(input_path,
+                                   nullptr,
+                                   nullptr,
+                                   nullptr,
+                                   nullptr,
+                                   vae_path,
+                                   output_path,
+                                   output_type,
+                                   tensor_type_rules,
+                                   convert_name,
+                                   0);
 }

src/core/backend_fit.cpp

@@ -0,0 +1,390 @@
+#include "backend_fit.h"
+
+#include <algorithm>
+#include <cctype>
+#include <cstdint>
+#include <utility>
+#include <vector>
+
+#include "core/ggml_extend_backend.h"
+#include "core/util.h"
+#include "ggml-backend.h"
+
+namespace sd::backend_fit {
+    namespace {
+
+        constexpr int64_t MiB = 1024ll * 1024;
+
+        enum class ComponentKind {
+            DIT         = 0,
+            VAE         = 1,
+            CONDITIONER = 2,
+        };
+
+        struct Component {
+            ComponentKind kind;
+            const char* name;
+            int64_t params_bytes  = 0;
+            int64_t reserve_bytes = 0;
+            bool splittable       = false;
+        };
+
+        struct Device {
+            ggml_backend_dev_t dev = nullptr;
+            std::string name;
+            std::string description;
+            int64_t free_bytes   = 0;
+            int64_t total_bytes  = 0;
+            int64_t budget_bytes = 0;
+        };
+
+        struct Decision {
+            ComponentKind kind;
+            bool on_cpu = false;
+            std::vector<size_t> device_idxs;
+        };
+
+        struct Plan {
+            bool valid      = false;
+            bool time_share = false;
+            std::vector<Decision> decisions;
+        };
+
+        bool classify_tensor(const std::string& name, ComponentKind& out) {
+            auto contains = [&](const char* s) { return name.find(s) != std::string::npos; };
+
+            if (contains("model.diffusion_model.") || contains("unet.")) {
+                out = ComponentKind::DIT;
+                return true;
+            }
+            if (contains("first_stage_model.") ||
+                name.rfind("vae.", 0) == 0 ||
+                name.rfind("tae.", 0) == 0) {
+                out = ComponentKind::VAE;
+                return true;
+            }
+            if (contains("text_encoders") ||
+                contains("cond_stage_model") ||
+                contains("te.text_model.") ||
+                contains("conditioner") ||
+                name.rfind("text_encoder.", 0) == 0 ||
+                name.rfind("text_embedding_projection.", 0) == 0 ||
+                contains(".aggregate_embed.")) {
+                out = ComponentKind::CONDITIONER;
+                return true;
+            }
+            return false;
+        }
+
+        std::vector<Component> estimate_components(ModelLoader& loader, ggml_type override_wtype) {
+            const auto& storage = loader.get_tensor_storage_map();
+
+            int64_t bytes[3] = {0, 0, 0};
+            for (const auto& [name, ts_const] : storage) {
+                TensorStorage ts = ts_const;
+                if (is_unused_tensor(ts.name)) {
+                    continue;
+                }
+                ComponentKind kind;
+                if (!classify_tensor(ts.name, kind)) {
+                    continue;
+                }
+                if (override_wtype != GGML_TYPE_COUNT &&
+                    loader.tensor_should_be_converted(ts, override_wtype)) {
+                    ts.type = override_wtype;
+                } else if (ts.expected_type != GGML_TYPE_COUNT && ts.expected_type != ts.type) {
+                    ts.type = ts.expected_type;
+                }
+                bytes[int(kind)] += (int64_t)ts.nbytes() + 64;
+            }
+
+            std::vector<Component> out;
+            out.push_back({ComponentKind::DIT, "DiT", bytes[int(ComponentKind::DIT)], 2048 * MiB, true});
+            out.push_back({ComponentKind::VAE, "VAE", bytes[int(ComponentKind::VAE)], 1024 * MiB, false});
+            out.push_back({ComponentKind::CONDITIONER, "Conditioner", bytes[int(ComponentKind::CONDITIONER)], 2048 * MiB, true});
+            return out;
+        }
+
+        std::vector<Device> enumerate_gpu_devices(const sd::ggml_graph_cut::MaxVramAssignment& budgets) {
+            std::vector<Device> out;
+            for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+                ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+                    continue;
+                }
+                Device d;
+                d.dev             = dev;
+                d.name            = ggml_backend_dev_name(dev);
+                d.description     = ggml_backend_dev_description(dev);
+                size_t free_bytes = 0, total_bytes = 0;
+                ggml_backend_dev_memory(dev, &free_bytes, &total_bytes);
+                d.free_bytes  = (int64_t)free_bytes;
+                d.total_bytes = (int64_t)total_bytes;
+
+                std::string budget_key = d.name;
+                std::transform(budget_key.begin(), budget_key.end(), budget_key.begin(),
+                               [](unsigned char c) { return (char)std::tolower(c); });
+                float gib = budgets.default_gib;
+                auto it   = budgets.backend_gib.find(budget_key);
+                if (it != budgets.backend_gib.end()) {
+                    gib = it->second;
+                }
+                if (gib > 0.f) {
+                    d.budget_bytes = std::min<int64_t>((int64_t)(gib * 1024.0 * 1024.0 * 1024.0), d.free_bytes);
+                } else if (gib < 0.f) {
+                    d.budget_bytes = d.free_bytes + (int64_t)(gib * 1024.0 * 1024.0 * 1024.0);
+                } else {
+                    d.budget_bytes = d.free_bytes - 512 * MiB;
+                }
+                d.budget_bytes = std::max<int64_t>(d.budget_bytes, 0);
+                out.push_back(d);
+            }
+            return out;
+        }
+
+        Plan compute_plan(const std::vector<Component>& components, const std::vector<Device>& devices) {
+            Plan plan;
+            if (devices.empty()) {
+                return plan;
+            }
+
+            std::vector<size_t> order(components.size());
+            for (size_t i = 0; i < order.size(); i++) {
+                order[i] = i;
+            }
+            std::sort(order.begin(), order.end(), [&](size_t a, size_t b) {
+                return components[a].params_bytes > components[b].params_bytes;
+            });
+
+            {
+                std::vector<int64_t> params_sum(devices.size(), 0);
+                std::vector<int64_t> max_reserve(devices.size(), 0);
+                std::vector<Decision> decisions(components.size());
+                bool ok = true;
+                for (size_t ci : order) {
+                    const Component& comp = components[ci];
+                    decisions[ci].kind    = comp.kind;
+                    if (comp.params_bytes == 0) {
+                        continue;
+                    }
+                    int best = -1;
+                    for (size_t di = 0; di < devices.size(); di++) {
+                        int64_t need = params_sum[di] + comp.params_bytes + std::max(max_reserve[di], comp.reserve_bytes);
+                        if (need <= devices[di].budget_bytes &&
+                            (best < 0 || devices[di].budget_bytes - params_sum[di] > devices[best].budget_bytes - params_sum[best])) {
+                            best = (int)di;
+                        }
+                    }
+                    if (best < 0) {
+                        ok = false;
+                        break;
+                    }
+                    params_sum[best] += comp.params_bytes;
+                    max_reserve[best] = std::max(max_reserve[best], comp.reserve_bytes);
+                    decisions[ci].device_idxs.push_back((size_t)best);
+                }
+                if (ok) {
+                    plan.valid      = true;
+                    plan.time_share = false;
+                    plan.decisions  = std::move(decisions);
+                    return plan;
+                }
+            }
+
+            plan.decisions.assign(components.size(), {});
+            for (size_t ci : order) {
+                const Component& comp = components[ci];
+                Decision& decision    = plan.decisions[ci];
+                decision.kind         = comp.kind;
+                if (comp.params_bytes == 0) {
+                    continue;
+                }
+                int best = -1;
+                for (size_t di = 0; di < devices.size(); di++) {
+                    if (comp.params_bytes + comp.reserve_bytes <= devices[di].budget_bytes &&
+                        (best < 0 || devices[di].budget_bytes > devices[best].budget_bytes)) {
+                        best = (int)di;
+                    }
+                }
+                if (best >= 0) {
+                    decision.device_idxs.push_back((size_t)best);
+                    continue;
+                }
+                if (comp.splittable && devices.size() > 1) {
+                    int64_t capacity = 0;
+                    for (const Device& d : devices) {
+                        capacity += std::max<int64_t>(d.budget_bytes - comp.reserve_bytes, 0);
+                    }
+                    if (comp.params_bytes <= capacity) {
+                        std::vector<size_t> idxs(devices.size());
+                        for (size_t i = 0; i < idxs.size(); i++) {
+                            idxs[i] = i;
+                        }
+                        std::sort(idxs.begin(), idxs.end(), [&](size_t a, size_t b) {
+                            return devices[a].budget_bytes > devices[b].budget_bytes;
+                        });
+                        decision.device_idxs = std::move(idxs);
+                        continue;
+                    }
+                }
+                decision.on_cpu = true;
+            }
+            plan.valid      = true;
+            plan.time_share = true;
+            return plan;
+        }
+
+        void print_plan(const Plan& plan,
+                        const std::vector<Component>& components,
+                        const std::vector<Device>& devices) {
+            LOG_INFO("auto-fit plan%s:", plan.time_share ? " (time-share: params load per phase and free after)" : "");
+            LOG_INFO("  devices:");
+            for (const Device& d : devices) {
+                LOG_INFO("    %-12s %-32s free %6lld MiB, budget %6lld MiB",
+                         d.name.c_str(), d.description.c_str(),
+                         (long long)(d.free_bytes / MiB), (long long)(d.budget_bytes / MiB));
+            }
+            LOG_INFO("  components:");
+            for (size_t ci = 0; ci < components.size(); ci++) {
+                const Component& comp    = components[ci];
+                const Decision& decision = plan.decisions[ci];
+                std::string target;
+                if (comp.params_bytes == 0) {
+                    target = "(not present)";
+                } else if (decision.on_cpu) {
+                    target = "CPU";
+                } else {
+                    for (size_t k = 0; k < decision.device_idxs.size(); k++) {
+                        if (k > 0) {
+                            target += " & ";
+                        }
+                        target += devices[decision.device_idxs[k]].name;
+                    }
+                    if (decision.device_idxs.size() > 1) {
+                        target += " (split)";
+                    }
+                }
+                LOG_INFO("    %-12s params %6lld MiB, compute reserve %5lld MiB -> %s",
+                         comp.name,
+                         (long long)(comp.params_bytes / MiB),
+                         (long long)(comp.reserve_bytes / MiB),
+                         target.c_str());
+            }
+        }
+
+        void append_assignment(std::string& spec, const char* key, const std::string& value) {
+            if (!spec.empty()) {
+                spec += ",";
+            }
+            spec += key;
+            spec += "=";
+            spec += value;
+        }
+
+        void append_component_decision(const std::vector<Component>& components,
+                                       const std::vector<Device>& devices,
+                                       const Plan& plan,
+                                       ComponentKind kind,
+                                       const char* module_key,
+                                       std::string& runtime_spec,
+                                       std::string& params_spec) {
+            for (size_t ci = 0; ci < components.size(); ci++) {
+                if (components[ci].kind != kind || components[ci].params_bytes == 0) {
+                    continue;
+                }
+                const Decision& decision = plan.decisions[ci];
+                if (decision.on_cpu) {
+                    append_assignment(runtime_spec, module_key, "cpu");
+                    return;
+                }
+                if (decision.device_idxs.empty()) {
+                    return;
+                }
+                std::string device_list;
+                for (size_t k = 0; k < decision.device_idxs.size(); k++) {
+                    if (k > 0) {
+                        device_list += "&";
+                    }
+                    device_list += devices[decision.device_idxs[k]].name;
+                }
+                append_assignment(runtime_spec, module_key, device_list);
+                if (plan.time_share) {
+                    append_assignment(params_spec, module_key, "disk");
+                }
+                return;
+            }
+        }
+
+    }  // namespace
+
+    bool derive_backend_specs(ModelLoader& loader,
+                              ggml_type override_wtype,
+                              sd::ggml_graph_cut::MaxVramAssignment& budgets,
+                              std::string& runtime_spec,
+                              std::string& params_spec) {
+        if (!runtime_spec.empty() || !params_spec.empty()) {
+            LOG_WARN("--auto-fit is enabled; ignoring --backend / --params-backend");
+        }
+
+        {
+            std::string error;
+            if (!budgets.canonicalize_backend_keys(&error)) {
+                LOG_ERROR("%s", error.c_str());
+                return false;
+            }
+        }
+
+        auto components = estimate_components(loader, override_wtype);
+        auto devices    = enumerate_gpu_devices(budgets);
+        auto plan       = compute_plan(components, devices);
+        if (!plan.valid) {
+            LOG_WARN("auto-fit: no usable GPU devices; using the default backend");
+            runtime_spec.clear();
+            params_spec.clear();
+            return true;
+        }
+
+        print_plan(plan, components, devices);
+
+        std::string derived_runtime_spec;
+        std::string derived_params_spec;
+        append_component_decision(components, devices, plan, ComponentKind::DIT, "diffusion", derived_runtime_spec, derived_params_spec);
+        append_component_decision(components, devices, plan, ComponentKind::CONDITIONER, "te", derived_runtime_spec, derived_params_spec);
+        append_component_decision(components, devices, plan, ComponentKind::VAE, "vae", derived_runtime_spec, derived_params_spec);
+
+        runtime_spec = std::move(derived_runtime_spec);
+        params_spec  = std::move(derived_params_spec);
+
+        LOG_INFO("auto-fit: --backend \"%s\"%s%s%s",
+                 runtime_spec.empty() ? "(default)" : runtime_spec.c_str(),
+                 params_spec.empty() ? "" : " --params-backend \"",
+                 params_spec.c_str(),
+                 params_spec.empty() ? "" : "\"");
+        return true;
+    }
+
+    bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params, bool prefer_temporal_tiling) {
+        if (prefer_temporal_tiling) {
+            if (tiling_params.temporal_tiling) {
+                return false;
+            }
+            tiling_params.temporal_tiling = true;
+        } else {
+            if (tiling_params.enabled) {
+                return false;
+            }
+            tiling_params.enabled = true;
+            if (tiling_params.tile_size_x <= 0) {
+                tiling_params.tile_size_x = 256;
+            }
+            if (tiling_params.tile_size_y <= 0) {
+                tiling_params.tile_size_y = 256;
+            }
+        }
+
+        LOG_WARN("auto-fit: VAE decode failed (likely out of memory); retrying with %s tiling",
+                 tiling_params.temporal_tiling ? "temporal" : "spatial");
+        return true;
+    }
+
+}  // namespace sd::backend_fit

src/core/backend_fit.h

@@ -0,0 +1,23 @@
+#ifndef __SD_BACKEND_FIT_H__
+#define __SD_BACKEND_FIT_H__
+
+#include <string>
+
+#include "core/ggml_graph_cut.h"
+#include "model_loader.h"
+#include "stable-diffusion.h"
+
+namespace sd::backend_fit {
+
+    bool derive_backend_specs(ModelLoader& loader,
+                              ggml_type override_wtype,
+                              sd::ggml_graph_cut::MaxVramAssignment& budgets,
+                              std::string& runtime_spec,
+                              std::string& params_spec);
+
+    bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params,
+                                         bool prefer_temporal_tiling);
+
+}  // namespace sd::backend_fit
+
+#endif  // __SD_BACKEND_FIT_H__

src/core/ggml_extend.hpp

@@ -391,7 +391,7 @@ __STATIC_INLINE__ uint8_t* ggml_tensor_to_sd_image(ggml_tensor* input, uint8_t*
     int64_t width    = input->ne[0];
     int64_t height   = input->ne[1];
     int64_t channels = input->ne[2];
-    GGML_ASSERT(channels == 3 && input->type == GGML_TYPE_F32);
+    GGML_ASSERT(input->type == GGML_TYPE_F32);
     if (image_data == nullptr) {
         image_data = (uint8_t*)malloc(width * height * channels);
     }
@@ -1038,6 +1038,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_linear(ggml_context* ctx,
 }
 
 __STATIC_INLINE__ ggml_tensor* ggml_ext_pad_ext(ggml_context* ctx,
+                                                ggml_backend_t backend,
                                                 ggml_tensor* x,
                                                 int lp0,
                                                 int rp0,
@@ -1063,7 +1064,17 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_pad_ext(ggml_context* ctx,
     }
 
     if (lp0 != 0 || rp0 != 0 || lp1 != 0 || rp1 != 0 || lp2 != 0 || rp2 != 0 || lp3 != 0 || rp3 != 0) {
-        x = ggml_pad_ext(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+        ggml_tensor* padded = ggml_pad_ext(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+        if (backend == nullptr || ggml_backend_supports_op(backend, padded)) {
+            x = padded;
+        } else {
+            // Some backends (e.g. Metal) only implement right-padding for
+            // GGML_OP_PAD (see #850): pad right by lp+rp instead, then roll
+            // the padding around to the left. shift < ne always holds because
+            // ne grew by lp+rp.
+            x = ggml_pad_ext(ctx, x, 0, lp0 + rp0, 0, lp1 + rp1, 0, lp2 + rp2, 0, lp3 + rp3);
+            x = ggml_roll(ctx, x, lp0, lp1, lp2, lp3);
+        }
     }
     return x;
 }
@@ -1076,7 +1087,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_pad(ggml_context* ctx,
                                             int p3          = 0,
                                             bool circular_x = false,
                                             bool circular_y = false) {
-    return ggml_ext_pad_ext(ctx, x, 0, p0, 0, p1, 0, p2, 0, p3, circular_x, circular_y);
+    return ggml_ext_pad_ext(ctx, nullptr, x, 0, p0, 0, p1, 0, p2, 0, p3, circular_x, circular_y);
 }
 
 // w: [OC，IC, KH, KW]
@@ -1105,7 +1116,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_2d(ggml_context* ctx,
     }
 
     if ((p0 != 0 || p1 != 0) && (circular_x || circular_y)) {
-        x  = ggml_ext_pad_ext(ctx, x, p0, p0, p1, p1, 0, 0, 0, 0, circular_x, circular_y);
+        x  = ggml_ext_pad_ext(ctx, nullptr, x, p0, p0, p1, p1, 0, 0, 0, 0, circular_x, circular_y);
         p0 = 0;
         p1 = 0;
     }
@@ -1130,6 +1141,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_2d(ggml_context* ctx,
 // b: [OC,]
 // result: [N*OC, OD, OH, OW]
 __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
+                                                ggml_backend_t backend,
                                                 ggml_tensor* x,
                                                 ggml_tensor* w,
                                                 ggml_tensor* b,
@@ -1159,7 +1171,21 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
         x          = ggml_cont(ctx, ggml_permute(ctx, x, 0, 1, 3, 2));
         x          = ggml_reshape_4d(ctx, x, im2col->ne[1], im2col->ne[2], OD, OC * N);
     } else {
-        x = ggml_conv_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2);
+        // ggml_conv_3d decomposes into GGML_OP_IM2COL_3D, which some backends
+        // (e.g. Metal, see #850) do not implement. Fall back to
+        // GGML_OP_CONV_3D on those backends.
+        bool im2col_3d_supported = true;
+        if (backend != nullptr) {
+            ggml_tensor* im2col = ggml_im2col_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, w->type);
+            im2col_3d_supported = ggml_backend_supports_op(backend, im2col);
+        }
+        if (im2col_3d_supported) {
+            x = ggml_conv_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2);
+        } else {
+            int64_t OC = w->ne[3] / IC;
+            int64_t N  = x->ne[3] / IC;
+            x          = ggml_conv_3d_direct(ctx, w, x, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
+        }
     }
 
     if (b != nullptr) {
@@ -1362,6 +1388,9 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         }
 
         auto out = ggml_flash_attn_ext(ctx, q_in, k_in, v_in, mask_in, scale / kv_scale, 0, 0);
+        if (!ggml_backend_supports_op(backend, out)) {
+            return nullptr;
+        }
         ggml_flash_attn_ext_set_prec(out, GGML_PREC_F32);
         if (kv_scale != 1.0f) {
             out = ggml_ext_scale(ctx, out, 1.0f / kv_scale);
@@ -1379,9 +1408,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
 
         if (can_use_flash_attn) {
             kqv = build_kqv(q, k, v, mask);
-            if (!ggml_backend_supports_op(backend, kqv)) {
-                kqv = nullptr;
-            } else {
+            if (kqv != nullptr) {
                 kqv = ggml_view_4d(ctx,
                                    kqv,
                                    d_head,
@@ -1719,6 +1746,11 @@ protected:
     bool stream_layers_enabled            = false;
     size_t observed_max_effective_budget_ = 0;
 
+    std::vector<ggml_backend_t> extra_runtime_backends;  // borrowed (SDBackendManager-owned)
+    ggml_backend_sched_t sched             = nullptr;    // owned, multi-device only
+    ggml_backend_t cpu_fallback_backend    = nullptr;    // owned, sched requires a trailing CPU backend
+    bool multi_device_eval_callback_warned = false;
+
     std::shared_ptr<WeightAdapter> weight_adapter = nullptr;
     std::weak_ptr<RunnerWeightManager> weight_manager;
     std::unordered_set<const ggml_tensor*> kept_compute_param_tensor_set;
@@ -1986,7 +2018,121 @@ protected:
         return true;
     }
 
+    // Pass explicit buffer types: synthesized defaults can make CUDA devices
+    // report supporting each other's buffers and skip a required copy.
+    bool ensure_sched(ggml_cgraph* gf) {
+        if (sched != nullptr) {
+            return true;
+        }
+        std::vector<ggml_backend_t> backends;
+        backends.reserve(extra_runtime_backends.size() + 2);
+        backends.push_back(runtime_backend);
+        for (ggml_backend_t backend : extra_runtime_backends) {
+            backends.push_back(backend);
+        }
+        if (cpu_fallback_backend == nullptr && !sd_backend_is_cpu(runtime_backend)) {
+            cpu_fallback_backend = sd_backend_cpu_init();
+        }
+        if (cpu_fallback_backend != nullptr) {
+            backends.push_back(cpu_fallback_backend);
+        }
+
+        std::vector<ggml_backend_buffer_type_t> bufts;
+        bufts.reserve(backends.size());
+        ggml_backend_dev_t main_dev = ggml_backend_get_device(runtime_backend);
+        for (ggml_backend_t backend : backends) {
+            ggml_backend_buffer_type_t buft = nullptr;
+            if (backend == cpu_fallback_backend && main_dev != nullptr) {
+                buft = ggml_backend_dev_host_buffer_type(main_dev);
+            }
+            if (buft == nullptr) {
+                buft = ggml_backend_get_default_buffer_type(backend);
+            }
+            bufts.push_back(buft);
+        }
+
+        size_t graph_size = MAX_GRAPH_SIZE;
+        if (gf != nullptr) {
+            graph_size = std::max<size_t>(graph_size, (size_t)ggml_graph_n_nodes(gf));
+        }
+        sched = ggml_backend_sched_new(backends.data(),
+                                       bufts.data(),
+                                       (int)backends.size(),
+                                       graph_size,
+                                       /*parallel=*/false,
+                                       /*op_offload=*/false);
+        if (sched == nullptr) {
+            LOG_ERROR("%s: failed to create backend sched", get_desc().c_str());
+            return false;
+        }
+        return true;
+    }
+
+    ggml_backend_t backend_for_weight(const ggml_tensor* tensor) const {
+        if (tensor == nullptr || tensor->buffer == nullptr) {
+            return nullptr;
+        }
+        if (ggml_backend_buffer_get_usage(tensor->buffer) != GGML_BACKEND_BUFFER_USAGE_WEIGHTS ||
+            ggml_backend_buffer_is_host(tensor->buffer)) {
+            return nullptr;
+        }
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(ggml_backend_buffer_get_type(tensor->buffer));
+        if (dev == nullptr) {
+            return nullptr;
+        }
+        if (ggml_backend_get_device(runtime_backend) == dev) {
+            return runtime_backend;
+        }
+        for (ggml_backend_t backend : extra_runtime_backends) {
+            if (ggml_backend_get_device(backend) == dev) {
+                return backend;
+            }
+        }
+        return nullptr;
+    }
+
+    // Weightless ops have no scheduler anchor, so pin them to the most recent
+    // weight device. Views must stay unpinned or cross-device copies can be
+    // skipped for their consumers.
+    void pin_multi_device_nodes(ggml_cgraph* gf) {
+        if (sched == nullptr || gf == nullptr) {
+            return;
+        }
+        ggml_backend_t current = runtime_backend;
+        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);
+            for (int s = 0; s < GGML_MAX_SRC; s++) {
+                ggml_backend_t weight_backend = backend_for_weight(node->src[s]);
+                if (weight_backend != nullptr) {
+                    current = weight_backend;
+                }
+            }
+            if (node->op == GGML_OP_NONE || node->op == GGML_OP_VIEW || node->op == GGML_OP_RESHAPE ||
+                node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE) {
+                continue;
+            }
+            if (ggml_backend_supports_op(current, node)) {
+                ggml_backend_sched_set_tensor_backend(sched, node, current);
+            }
+        }
+    }
+
+    bool is_multi_device() const {
+        return !extra_runtime_backends.empty();
+    }
+
     bool alloc_compute_buffer(ggml_cgraph* gf) {
+        if (is_multi_device()) {
+            // The sched replaces the gallocr. Do NOT ggml_backend_sched_reserve
+            // the graph here: reserve runs split_graph, which rewires the
+            // graph's src pointers to sched-internal copy tensors, and the
+            // later ggml_backend_sched_alloc_graph would split the already
+            // rewired graph, silently corrupting every cross-backend input. A
+            // graph must be split at most once; the alloc in execute_graph
+            // performs the real allocation.
+            return ensure_sched(gf);
+        }
         if (compute_allocr != nullptr) {
             return true;
         }
@@ -2202,12 +2348,14 @@ protected:
                plan.valid &&
                max_graph_vram_bytes > 0 &&
                plan.segments.size() > 1 &&
-               !sd_backend_is_cpu(runtime_backend);
+               !sd_backend_is_cpu(runtime_backend) &&
+               !is_multi_device();
     }
 
     bool can_attempt_graph_cut_segmented_compute() const {
         return max_graph_vram_bytes > 0 &&
-               !sd_backend_is_cpu(runtime_backend);
+               !sd_backend_is_cpu(runtime_backend) &&
+               !is_multi_device();
     }
 
     bool resolve_graph_cut_plan(ggml_cgraph* gf,
@@ -2463,7 +2611,14 @@ protected:
         };
         ComputeBufferGuard compute_buffer_guard(this, free_compute_buffer);
 
-        if (!ggml_gallocr_alloc_graph(compute_allocr, gf)) {
+        if (is_multi_device()) {
+            ggml_backend_sched_reset(sched);
+            pin_multi_device_nodes(gf);  // reset clears the pins; re-apply before alloc
+            if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+                LOG_ERROR("%s sched alloc compute graph failed", get_desc().c_str());
+                return std::nullopt;
+            }
+        } else if (!ggml_gallocr_alloc_graph(compute_allocr, gf)) {
             LOG_ERROR("%s alloc compute graph failed", get_desc().c_str());
             return std::nullopt;
         }
@@ -2472,8 +2627,27 @@ protected:
         if (sd_backend_is_cpu(runtime_backend)) {
             sd_backend_cpu_set_n_threads(runtime_backend, n_threads);
         }
+        if (cpu_fallback_backend != nullptr) {
+            sd_backend_cpu_set_n_threads(cpu_fallback_backend, n_threads);
+        }
 
-        ggml_status status = ggml_backend_graph_compute(runtime_backend, gf);
+        ggml_status status;
+        if (is_multi_device()) {
+            if (sd_get_backend_eval_callback() != nullptr && !multi_device_eval_callback_warned) {
+                LOG_WARN("%s: eval callback is not supported with multiple runtime backends; ignoring",
+                         get_desc().c_str());
+                multi_device_eval_callback_warned = true;
+            }
+            status = ggml_backend_sched_graph_compute(sched, gf);
+            if (status == GGML_STATUS_SUCCESS) {
+                ggml_backend_sched_synchronize(sched);
+            }
+        } else {
+            status = sd_backend_graph_compute_with_eval_callback(runtime_backend,
+                                                                 gf,
+                                                                 sd_get_backend_eval_callback(),
+                                                                 sd_get_backend_eval_callback_data());
+        }
         if (status != GGML_STATUS_SUCCESS) {
             LOG_ERROR("%s compute failed: %s", get_desc().c_str(), ggml_status_to_string(status));
             return std::nullopt;
@@ -2650,6 +2824,10 @@ public:
         free_params_ctx();
         free_compute_ctx();
         free_cache_ctx_and_buffer();
+        if (cpu_fallback_backend != nullptr) {
+            ggml_backend_free(cpu_fallback_backend);
+            cpu_fallback_backend = nullptr;
+        }
     }
 
     virtual GGMLRunnerContext get_context() {
@@ -2690,10 +2868,20 @@ public:
             ggml_gallocr_free(compute_allocr);
             compute_allocr = nullptr;
         }
+        if (sched != nullptr) {
+            ggml_backend_sched_free(sched);
+            sched = nullptr;
+        }
     }
 
     // do copy after alloc graph
     void set_backend_tensor_data(ggml_tensor* tensor, const void* data) {
+        if (is_multi_device()) {
+            // The sched only assigns a backend (and thus a buffer) to tensors
+            // that participate in the graph; flag standalone data tensors as
+            // inputs so they get one.
+            ggml_set_input(tensor);
+        }
         backend_tensor_data_map[tensor] = data;
     }
 
@@ -2829,8 +3017,31 @@ public:
     }
 
     void set_stream_layers_enabled(bool enabled) {
+        if (enabled && is_multi_device()) {
+            LOG_WARN("%s: --stream-layers is not supported with multiple runtime backends; ignoring",
+                     get_desc().c_str());
+            return;
+        }
         stream_layers_enabled = enabled;
     }
+
+    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) {
+        extra_runtime_backends.clear();
+        for (ggml_backend_t backend : backends) {
+            if (backend == nullptr || backend == runtime_backend) {
+                continue;
+            }
+            if (std::find(extra_runtime_backends.begin(), extra_runtime_backends.end(), backend) ==
+                extra_runtime_backends.end()) {
+                extra_runtime_backends.push_back(backend);
+            }
+        }
+        if (is_multi_device() && stream_layers_enabled) {
+            LOG_WARN("%s: --stream-layers is not supported with multiple runtime backends; ignoring",
+                     get_desc().c_str());
+            stream_layers_enabled = false;
+        }
+    }
 };
 
 class GGMLBlock {
@@ -3365,7 +3576,7 @@ public:
                 b = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, b, prefix + "bias");
             }
         }
-        return ggml_ext_conv_3d(ctx->ggml_ctx, x, w, b, in_channels,
+        return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
                                 std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                 std::get<2>(padding), std::get<1>(padding), std::get<0>(padding),
                                 std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),

src/core/ggml_extend_backend.cpp

@@ -9,6 +9,7 @@
 #include <vector>
 
 #include "core/util.h"
+#include "ggml/src/ggml-impl.h"
 #include "stable-diffusion.h"
 
 static std::string trim_copy(const std::string& value) {
@@ -110,7 +111,67 @@ static std::string resolve_first_device_by_type(enum ggml_backend_dev_type type)
     if (dev == nullptr) {
         return "";
     }
-    return ggml_backend_dev_name(dev);
+    const char* dev_name = ggml_backend_dev_name(dev);
+    if (dev_name != nullptr && dev_name[0] != '\0') {
+        return dev_name;
+    }
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    const char* reg_name   = reg != nullptr ? ggml_backend_reg_name(reg) : nullptr;
+    return reg_name != nullptr ? reg_name : "";
+}
+
+static ggml_backend_dev_t resolve_first_device_by_registry_name(const std::string& name) {
+    std::string lower = lower_copy(trim_copy(name));
+    if (lower == "metal") {
+        lower = "mtl";
+    }
+    if (lower.empty()) {
+        return nullptr;
+    }
+
+    const size_t device_count = ggml_backend_dev_count();
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+        if (reg == nullptr) {
+            continue;
+        }
+        const char* reg_name = ggml_backend_reg_name(reg);
+        if (reg_name != nullptr && lower_copy(reg_name) == lower) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+static ggml_backend_dev_t resolve_device_by_name(const std::string& name) {
+    const std::string lower = lower_copy(trim_copy(name));
+    if (lower.empty()) {
+        return nullptr;
+    }
+
+    const size_t device_count = ggml_backend_dev_count();
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        const char* dev_name   = ggml_backend_dev_name(dev);
+        if (dev_name != nullptr && lower_copy(dev_name) == lower) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+static std::string backend_device_name(ggml_backend_dev_t dev) {
+    if (dev == nullptr) {
+        return "";
+    }
+    const char* name = ggml_backend_dev_name(dev);
+    if (name != nullptr && name[0] != '\0') {
+        return name;
+    }
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    const char* reg_name   = reg != nullptr ? ggml_backend_reg_name(reg) : nullptr;
+    return reg_name != nullptr ? reg_name : "";
 }
 
 static ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
@@ -296,6 +357,10 @@ std::string sd_backend_resolve_name(const std::string& name) {
         return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
     }
 
+    if (ggml_backend_dev_t dev = resolve_first_device_by_registry_name(requested)) {
+        return backend_device_name(dev);
+    }
+
     const size_t device_count = ggml_backend_dev_count();
     for (size_t i = 0; i < device_count; ++i) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
@@ -328,7 +393,20 @@ static ggml_backend_t init_named_backend(const std::string& name) {
         return ggml_backend_init_best();
     }
 
+    if (ggml_backend_dev_t dev = resolve_device_by_name(name)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
+    if (ggml_backend_dev_t dev = resolve_first_device_by_registry_name(name)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
+
     std::string resolved = sd_backend_resolve_name(name);
+    if (ggml_backend_dev_t dev = resolve_device_by_name(resolved)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
+    if (ggml_backend_dev_t dev = resolve_first_device_by_registry_name(resolved)) {
+        return ggml_backend_dev_init(dev, nullptr);
+    }
     if (resolved.empty()) {
         return nullptr;
     }
@@ -364,6 +442,68 @@ bool sd_backend_cpu_set_n_threads(ggml_backend_t backend, int n_threads) {
     return false;
 }
 
+static ggml_cgraph sd_ggml_graph_view(ggml_cgraph* cgraph0, int i0, int i1) {
+    ggml_cgraph cgraph = {
+        /*.size             =*/0,
+        /*.n_nodes          =*/i1 - i0,
+        /*.n_leafs          =*/0,
+        /*.nodes            =*/cgraph0->nodes + i0,
+        /*.grads            =*/nullptr,
+        /*.grad_accs        =*/nullptr,
+        /*.leafs            =*/nullptr,
+        /*.use_counts       =*/cgraph0->use_counts,
+        /*.visited_hash_set =*/cgraph0->visited_hash_set,
+        /*.order            =*/cgraph0->order,
+        /*.uid              =*/0,
+    };
+    return cgraph;
+}
+
+ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,
+                                                        ggml_cgraph* gf,
+                                                        sd_graph_eval_callback_t callback_eval,
+                                                        void* callback_eval_user_data) {
+    if (callback_eval == nullptr) {
+        return ggml_backend_graph_compute(backend, gf);
+    }
+
+    ggml_status status = GGML_STATUS_SUCCESS;
+    const int n_nodes  = ggml_graph_n_nodes(gf);
+    bool stopped       = false;
+
+    for (int j0 = 0; j0 < n_nodes; ++j0) {
+        ggml_tensor* t = ggml_graph_node(gf, j0);
+        bool need      = callback_eval(t, true, callback_eval_user_data);
+        int j1         = j0;
+
+        while (!need && j1 < n_nodes - 1) {
+            t    = ggml_graph_node(gf, ++j1);
+            need = callback_eval(t, true, callback_eval_user_data);
+        }
+
+        ggml_cgraph gv = sd_ggml_graph_view(gf, j0, j1 + 1);
+        status         = ggml_backend_graph_compute_async(backend, &gv);
+        if (status != GGML_STATUS_SUCCESS) {
+            break;
+        }
+
+        ggml_backend_synchronize(backend);
+
+        if (need && !callback_eval(t, false, callback_eval_user_data)) {
+            stopped = true;
+            break;
+        }
+
+        j0 = j1;
+    }
+
+    ggml_backend_synchronize(backend);
+    if (stopped && status == GGML_STATUS_SUCCESS) {
+        status = GGML_STATUS_ABORTED;
+    }
+    return status;
+}
+
 const char* sd_get_system_info() {
     static std::string cache_info = []() -> std::string {
         ggml_backend_load_all_once();
@@ -525,12 +665,52 @@ SDBackendManager::~SDBackendManager() {
 
 void SDBackendManager::reset() {
     backends_.clear();
-    runtime_assignment_ = {};
-    params_assignment_  = {};
+    runtime_assignment_    = {};
+    params_assignment_     = {};
+    split_mode_assignment_ = {};
+}
+
+static std::vector<std::string> split_device_list(const std::string& value) {
+    std::vector<std::string> names;
+    for (const std::string& raw : split_copy(value, '&')) {
+        const std::string name = trim_copy(raw);
+        if (!name.empty()) {
+            names.push_back(name);
+        }
+    }
+    return names;
+}
+
+static std::string primary_device_name(const std::string& value) {
+    std::vector<std::string> names = split_device_list(value);
+    return names.empty() ? std::string() : names.front();
 }
 
 ggml_backend_t SDBackendManager::runtime_backend(SDBackendModule module) {
-    return init_cached_backend(runtime_assignment_.get(module));
+    return init_cached_backend(primary_device_name(runtime_assignment_.get(module)));
+}
+
+std::vector<ggml_backend_t> SDBackendManager::runtime_backends(SDBackendModule module) {
+    std::vector<ggml_backend_t> backends;
+    for (const std::string& name : split_device_list(runtime_assignment_.get(module))) {
+        ggml_backend_t backend = init_cached_backend(name);
+        if (backend == nullptr) {
+            LOG_ERROR("failed to initialize backend '%s' for module %s",
+                      name.c_str(),
+                      sd_backend_module_name(module));
+            continue;
+        }
+        if (std::find(backends.begin(), backends.end(), backend) == backends.end()) {
+            backends.push_back(backend);
+        }
+    }
+    if (backends.empty()) {
+        ggml_backend_t backend = runtime_backend(module);
+        if (backend != nullptr) {
+            backends.push_back(backend);
+        }
+    }
+    return backends;
 }
 
 ggml_backend_t SDBackendManager::params_backend(SDBackendModule module) {
@@ -556,6 +736,10 @@ bool SDBackendManager::params_backend_is_disk(SDBackendModule module) const {
     return is_disk_backend_token(params_assignment_.get(module));
 }
 
+bool SDBackendManager::params_backend_follows_runtime(SDBackendModule module) const {
+    return params_assignment_.get(module).empty();
+}
+
 bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule module) {
     ggml_backend_t backend = runtime_backend(module);
     if (backend == nullptr) {
@@ -575,6 +759,7 @@ bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule modu
 
 bool SDBackendManager::init(const char* backend_spec,
                             const char* params_backend_spec,
+                            const char* split_mode_spec,
                             std::string* error) {
     reset();
 
@@ -584,12 +769,53 @@ bool SDBackendManager::init(const char* backend_spec,
     if (!sd_parse_backend_assignment(SAFE_STR(params_backend_spec), &params_assignment_, error)) {
         return false;
     }
+    if (!sd_parse_backend_assignment(SAFE_STR(split_mode_spec), &split_mode_assignment_, error)) {
+        return false;
+    }
 
     return validate(error);
 }
 
+SDSplitMode SDBackendManager::split_mode(SDBackendModule module) const {
+    return lower_copy(trim_copy(split_mode_assignment_.get(module))) == "row" ? SDSplitMode::ROW
+                                                                              : SDSplitMode::LAYER;
+}
+
+ggml_backend_buffer_type_t SDBackendManager::split_buffer_type(ggml_backend_t backend,
+                                                               const std::vector<float>& tensor_split) {
+    if (backend == nullptr) {
+        return nullptr;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (dev == nullptr) {
+        return nullptr;
+    }
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    if (reg == nullptr) {
+        return nullptr;
+    }
+    auto fn = (ggml_backend_split_buffer_type_t)ggml_backend_reg_get_proc_address(reg, "ggml_backend_split_buffer_type");
+    if (fn == nullptr) {
+        return nullptr;
+    }
+    int main_device        = -1;
+    const size_t dev_count = ggml_backend_reg_dev_count(reg);
+    for (size_t i = 0; i < dev_count; ++i) {
+        if (ggml_backend_reg_dev_get(reg, i) == dev) {
+            main_device = (int)i;
+            break;
+        }
+    }
+    if (main_device < 0) {
+        return nullptr;
+    }
+    std::vector<float> padded_split(std::max<size_t>(tensor_split.size(), 64), 0.0f);
+    std::copy(tensor_split.begin(), tensor_split.end(), padded_split.begin());
+    return fn(main_device, padded_split.data());
+}
+
 bool SDBackendManager::validate(std::string* error) const {
-    auto validate_runtime_name = [&](const std::string& name) -> bool {
+    auto validate_single_runtime_name = [&](const std::string& name) -> bool {
         if (is_default_backend_token(name)) {
             return true;
         }
@@ -599,7 +825,7 @@ bool SDBackendManager::validate(std::string* error) const {
             }
             return false;
         }
-        if (!sd_backend_resolve_name(name).empty()) {
+        if (!sd_backend_resolve_name(name).empty() || resolve_first_device_by_registry_name(name) != nullptr) {
             return true;
         }
         if (error != nullptr) {
@@ -607,15 +833,56 @@ bool SDBackendManager::validate(std::string* error) const {
         }
         return false;
     };
+    auto validate_runtime_name = [&](const std::string& name) -> bool {
+        if (name.find('&') == std::string::npos) {
+            return validate_single_runtime_name(name);
+        }
+        std::vector<std::string> names = split_device_list(name);
+        if (names.empty()) {
+            if (error != nullptr) {
+                *error = "invalid backend device list '" + name + "'";
+            }
+            return false;
+        }
+        for (const std::string& entry : names) {
+            if (is_default_backend_token(entry)) {
+                if (error != nullptr) {
+                    *error = "default backend token is not allowed in a device list '" + name + "'";
+                }
+                return false;
+            }
+            if (!validate_single_runtime_name(entry)) {
+                return false;
+            }
+        }
+        return true;
+    };
     auto validate_params_name = [&](const std::string& name) -> bool {
         if (is_disk_backend_token(name)) {
             return true;
         }
-        return validate_runtime_name(name);
+        if (name.find('&') != std::string::npos) {
+            if (error != nullptr) {
+                *error = "params_backend does not accept device lists ('" + name + "')";
+            }
+            return false;
+        }
+        return validate_single_runtime_name(name);
+    };
+    auto validate_split_mode_name = [&](const std::string& name) -> bool {
+        const std::string lower = lower_copy(trim_copy(name));
+        if (lower.empty() || lower == "layer" || lower == "row") {
+            return true;
+        }
+        if (error != nullptr) {
+            *error = "invalid split mode '" + name + "' (expected layer or row)";
+        }
+        return false;
     };
 
     if (!validate_runtime_name(runtime_assignment_.default_name) ||
-        !validate_params_name(params_assignment_.default_name)) {
+        !validate_params_name(params_assignment_.default_name) ||
+        !validate_split_mode_name(split_mode_assignment_.default_name)) {
         return false;
     }
     for (const auto& kv : runtime_assignment_.module_names) {
@@ -628,6 +895,11 @@ bool SDBackendManager::validate(std::string* error) const {
             return false;
         }
     }
+    for (const auto& kv : split_mode_assignment_.module_names) {
+        if (!validate_split_mode_name(kv.second)) {
+            return false;
+        }
+    }
     return true;
 }
 

src/core/ggml_extend_backend.h

@@ -6,9 +6,11 @@
 #include <memory>
 #include <string>
 #include <unordered_map>
+#include <vector>
 
 #include "ggml-backend.h"
 #include "ggml.h"
+#include "stable-diffusion.h"
 
 enum class SDBackendModule {
     DIFFUSION,
@@ -36,10 +38,16 @@ struct SDBackendHandleDeleter {
 
 using SDBackendHandle = std::unique_ptr<struct ggml_backend, SDBackendHandleDeleter>;
 
+enum class SDSplitMode {
+    LAYER,
+    ROW,
+};
+
 class SDBackendManager {
 private:
     SDBackendAssignment runtime_assignment_;
     SDBackendAssignment params_assignment_;
+    SDBackendAssignment split_mode_assignment_;
     std::unordered_map<std::string, SDBackendHandle> backends_;
 
 public:
@@ -51,15 +59,23 @@ public:
 
     bool init(const char* backend_spec,
               const char* params_backend_spec,
+              const char* split_mode_spec,
               std::string* error);
     void reset();
 
     ggml_backend_t runtime_backend(SDBackendModule module);
     ggml_backend_t params_backend(SDBackendModule module);
 
+    std::vector<ggml_backend_t> runtime_backends(SDBackendModule module);
+
+    SDSplitMode split_mode(SDBackendModule module) const;
+    ggml_backend_buffer_type_t split_buffer_type(ggml_backend_t backend,
+                                                 const std::vector<float>& tensor_split);
+
     bool runtime_backend_is_cpu(SDBackendModule module);
     bool params_backend_is_cpu(SDBackendModule module);
     bool params_backend_is_disk(SDBackendModule module) const;
+    bool params_backend_follows_runtime(SDBackendModule module) const;
     bool runtime_backend_supports_host_buffer(SDBackendModule module);
 
 private:
@@ -71,6 +87,10 @@ bool sd_backend_is(ggml_backend_t backend, const std::string& name);
 bool sd_backend_is_cpu(ggml_backend_t backend);
 ggml_backend_t sd_backend_cpu_init();
 bool sd_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,
+                                                        ggml_cgraph* gf,
+                                                        sd_graph_eval_callback_t callback_eval,
+                                                        void* callback_eval_user_data);
 std::string sd_backend_resolve_name(const std::string& name);
 const char* sd_backend_module_name(SDBackendModule module);
 void ggml_ext_im_set_f32_1d(const struct ggml_tensor* tensor, int i, float value);

src/core/layer_split_partition.cpp

@@ -0,0 +1,221 @@
+#include "core/layer_split_partition.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+
+#include "core/util.h"
+
+namespace sd {
+
+    static bool layer_split_path_segment_starts_at(const std::string& name, size_t pos) {
+        return pos == 0 || name[pos - 1] == '.';
+    }
+
+    static bool layer_split_has_path_segment(const std::string& name, const char* segment) {
+        size_t pos = name.find(segment);
+        while (pos != std::string::npos) {
+            if (layer_split_path_segment_starts_at(name, pos)) {
+                return true;
+            }
+            pos = name.find(segment, pos + 1);
+        }
+        return false;
+    }
+
+    int layer_split_tensor_block_index(const std::string& name) {
+        static const char* unet_block_segments[] = {"input_blocks.", "output_blocks.", "middle_block.",
+                                                    "down_blocks.", "up_blocks.", "mid_block."};
+        for (const char* segment : unet_block_segments) {
+            if (layer_split_has_path_segment(name, segment)) {
+                return -1;
+            }
+        }
+
+        static const char* block_keywords[] = {"transformer_blocks.", "joint_blocks.", "double_blocks.",
+                                               "single_blocks.", "blocks.", "block.", "layers."};
+        for (const char* keyword : block_keywords) {
+            size_t pos = name.find(keyword);
+            while (pos != std::string::npos) {
+                if (!layer_split_path_segment_starts_at(name, pos)) {
+                    pos = name.find(keyword, pos + 1);
+                    continue;
+                }
+                pos += std::strlen(keyword);
+                size_t end = pos;
+                while (end < name.size() && name[end] >= '0' && name[end] <= '9') {
+                    end++;
+                }
+                if (end > pos && (end == name.size() || name[end] == '.')) {
+                    return std::atoi(name.substr(pos, end - pos).c_str());
+                }
+                break;
+            }
+        }
+        return -1;
+    }
+
+    std::string layer_split_backend_device_display_name(ggml_backend_t backend) {
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        const char* name       = dev != nullptr ? ggml_backend_dev_name(dev) : ggml_backend_name(backend);
+        return name != nullptr ? name : "unknown";
+    }
+
+    static bool layer_split_backend_supports_tensor(ggml_backend_t backend, const ggml_tensor* tensor) {
+        return backend != nullptr && tensor != nullptr && ggml_backend_supports_op(backend, tensor);
+    }
+
+    static size_t layer_split_supported_target(const std::string& desc,
+                                               const std::string& tensor_name,
+                                               const ggml_tensor* tensor,
+                                               const std::vector<ggml_backend_t>& backends,
+                                               size_t preferred) {
+        if (tensor == nullptr || backends.empty()) {
+            return preferred;
+        }
+        size_t preferred_safe = std::min(preferred, backends.size() - 1);
+        if (layer_split_backend_supports_tensor(backends[preferred_safe], tensor)) {
+            return preferred_safe;
+        }
+        for (size_t i = 0; i < backends.size(); i++) {
+            if (layer_split_backend_supports_tensor(backends[i], tensor)) {
+                LOG_WARN("%s layer split: moving tensor '%s' from %s to %s because the preferred backend cannot run op=%s type=%s nbytes=%.2f MB",
+                         desc.c_str(),
+                         tensor_name.c_str(),
+                         layer_split_backend_device_display_name(backends[preferred_safe]).c_str(),
+                         layer_split_backend_device_display_name(backends[i]).c_str(),
+                         ggml_op_name(tensor->op),
+                         ggml_type_name(tensor->type),
+                         ggml_nbytes(tensor) / (1024.0 * 1024.0));
+                return i;
+            }
+        }
+        LOG_WARN("%s layer split: tensor '%s' is not supported by any split backend: op=%s type=%s nbytes=%.2f MB",
+                 desc.c_str(),
+                 tensor_name.c_str(),
+                 ggml_op_name(tensor->op),
+                 ggml_type_name(tensor->type),
+                 ggml_nbytes(tensor) / (1024.0 * 1024.0));
+        return preferred_safe;
+    }
+
+    std::vector<std::map<std::string, ggml_tensor*>> partition_layer_split_tensors(
+        const std::string& desc,
+        const std::map<std::string, ggml_tensor*>& tensors,
+        const std::map<std::string, ggml_tensor*>& split_tensors,
+        const std::vector<ggml_backend_t>& backends) {
+        std::vector<std::map<std::string, ggml_tensor*>> partitions(backends.size());
+        if (backends.empty()) {
+            LOG_WARN("%s: no backend available for a layer split", desc.c_str());
+            return partitions;
+        }
+
+        std::map<int, int64_t> block_bytes;
+        std::map<std::string, size_t> non_block_targets;
+        std::vector<int64_t> other_bytes_by_backend(backends.size(), 0);
+        int64_t total_block_bytes = 0;
+        int64_t total_other_bytes = 0;
+        int n_blocks              = 0;
+        for (const auto& kv : tensors) {
+            int64_t bytes = (int64_t)ggml_nbytes(kv.second);
+            int idx       = split_tensors.count(kv.first) != 0 ? layer_split_tensor_block_index(kv.first) : -1;
+            if (idx >= 0) {
+                block_bytes[idx] += bytes;
+                total_block_bytes += bytes;
+                n_blocks = std::max(n_blocks, idx + 1);
+            } else {
+                size_t target               = layer_split_supported_target(desc, kv.first, kv.second, backends, 0);
+                non_block_targets[kv.first] = target;
+                other_bytes_by_backend[target] += bytes;
+                total_other_bytes += bytes;
+            }
+        }
+        if (n_blocks == 0) {
+            LOG_WARN("%s: no transformer blocks found for a layer split; keeping tensors on compatible backends starting from %s",
+                     desc.c_str(),
+                     layer_split_backend_device_display_name(backends[0]).c_str());
+            for (const auto& kv : tensors) {
+                size_t target  = 0;
+                auto target_it = non_block_targets.find(kv.first);
+                if (target_it != non_block_targets.end()) {
+                    target = target_it->second;
+                }
+                partitions[target][kv.first] = kv.second;
+            }
+            return partitions;
+        }
+
+        // Reserve compute headroom and subtract each device's actual non-block
+        // bytes from its block budget.
+        constexpr int64_t compute_headroom_bytes = 2ll * 1024 * 1024 * 1024;
+        std::vector<double> device_weights(backends.size(), 1.0);
+        double weight_sum = 0.0;
+        for (size_t i = 0; i < backends.size(); i++) {
+            ggml_backend_dev_t dev = ggml_backend_get_device(backends[i]);
+            size_t free_bytes = 0, total_bytes = 0;
+            if (dev != nullptr) {
+                ggml_backend_dev_memory(dev, &free_bytes, &total_bytes);
+            }
+            // Keep a small share even for tight devices instead of dropping them.
+            int64_t usable_bytes = std::max<int64_t>((int64_t)free_bytes - compute_headroom_bytes,
+                                                     (int64_t)free_bytes / 8);
+            device_weights[i]    = usable_bytes > 0 ? (double)usable_bytes : 1.0;
+            weight_sum += device_weights[i];
+        }
+
+        std::vector<int64_t> block_budgets(backends.size(), 0);
+        const int64_t total_bytes = total_block_bytes + total_other_bytes;
+        for (size_t i = 0; i < backends.size(); i++) {
+            int64_t budget   = (int64_t)((double)total_bytes * device_weights[i] / weight_sum);
+            budget           = std::max<int64_t>(budget - other_bytes_by_backend[i], 0);
+            block_budgets[i] = budget;
+        }
+
+        std::vector<int> boundaries(backends.size(), n_blocks);
+        size_t current = 0;
+        int64_t used   = 0;
+        for (int b = 0; b < n_blocks; b++) {
+            int64_t bytes = block_bytes.count(b) != 0 ? block_bytes[b] : 0;
+            if (current + 1 < backends.size() && used > 0 && used + bytes > block_budgets[current]) {
+                boundaries[current] = b;
+                current++;
+                used = 0;
+            }
+            used += bytes;
+        }
+
+        for (const auto& kv : tensors) {
+            size_t target = 0;
+            int idx       = split_tensors.count(kv.first) != 0 ? layer_split_tensor_block_index(kv.first) : -1;
+            if (idx >= 0) {
+                while (target < boundaries.size() && idx >= boundaries[target]) {
+                    target++;
+                }
+                target = std::min(target, backends.size() - 1);
+                target = layer_split_supported_target(desc, kv.first, kv.second, backends, target);
+            } else {
+                auto target_it = non_block_targets.find(kv.first);
+                if (target_it != non_block_targets.end()) {
+                    target = target_it->second;
+                }
+            }
+            partitions[target][kv.first] = kv.second;
+        }
+
+        int range_start = 0;
+        for (size_t i = 0; i < backends.size(); i++) {
+            int range_end                = boundaries[i];
+            const char* non_block_suffix = other_bytes_by_backend[i] > 0 ? " + non-block tensors" : "";
+            LOG_INFO("%s layer split: %s <- blocks [%d, %d)%s",
+                     desc.c_str(),
+                     layer_split_backend_device_display_name(backends[i]).c_str(),
+                     range_start,
+                     range_end,
+                     non_block_suffix);
+            range_start = range_end;
+        }
+        return partitions;
+    }
+
+}  // namespace sd

src/core/layer_split_partition.h

@@ -0,0 +1,24 @@
+#ifndef __SD_CORE_LAYER_SPLIT_PARTITION_H__
+#define __SD_CORE_LAYER_SPLIT_PARTITION_H__
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+namespace sd {
+
+    std::string layer_split_backend_device_display_name(ggml_backend_t backend);
+    int layer_split_tensor_block_index(const std::string& name);
+
+    std::vector<std::map<std::string, ggml_tensor*>> partition_layer_split_tensors(
+        const std::string& desc,
+        const std::map<std::string, ggml_tensor*>& tensors,
+        const std::map<std::string, ggml_tensor*>& split_tensors,
+        const std::vector<ggml_backend_t>& backends);
+
+}  // namespace sd
+
+#endif  // __SD_CORE_LAYER_SPLIT_PARTITION_H__

src/core/util.cpp

@@ -4,6 +4,8 @@
 #include <cmath>
 #include <codecvt>
 #include <cstdarg>
+#include <cstdlib>
+#include <cstring>
 #include <exception>
 #include <fstream>
 #include <locale>
@@ -25,6 +27,7 @@
 #include <unistd.h>
 #endif
 
+#include "ggml-backend.h"
 #include "ggml.h"
 #include "stable-diffusion.h"
 
@@ -346,6 +349,9 @@ int sd_preview_interval              = 1;
 bool sd_preview_denoised             = true;
 bool sd_preview_noisy                = false;
 
+static sd_graph_eval_callback_t sd_backend_eval_cb = nullptr;
+static void* sd_backend_eval_cb_data               = nullptr;
+
 std::u32string utf8_to_utf32(const std::string& utf8_str) {
     std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
     return converter.from_bytes(utf8_str);
@@ -629,6 +635,11 @@ void sd_set_preview_callback(sd_preview_cb_t cb, preview_t mode, int interval, b
     sd_preview_noisy    = noisy;
 }
 
+void sd_set_backend_eval_callback(sd_graph_eval_callback_t cb, void* data) {
+    sd_backend_eval_cb      = cb;
+    sd_backend_eval_cb_data = data;
+}
+
 sd_preview_cb_t sd_get_preview_callback() {
     return sd_preview_cb;
 }
@@ -649,6 +660,14 @@ bool sd_should_preview_noisy() {
     return sd_preview_noisy;
 }
 
+sd_graph_eval_callback_t sd_get_backend_eval_callback() {
+    return sd_backend_eval_cb;
+}
+
+void* sd_get_backend_eval_callback_data() {
+    return sd_backend_eval_cb_data;
+}
+
 sd_progress_cb_t sd_get_progress_callback() {
     return sd_progress_cb;
 }
@@ -981,3 +1000,26 @@ std::vector<std::pair<std::string, float>> split_quotation_attention(
     }
     return result;
 }
+
+size_t sd_list_devices(char* buffer, size_t buffer_size) {
+    if (ggml_backend_dev_count() == 0) {
+        // dynamic-backend builds discover their backend modules at runtime
+        ggml_backend_load_all();
+    }
+
+    std::ostringstream oss;
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        const char* name       = ggml_backend_dev_name(dev);
+        const char* desc       = ggml_backend_dev_description(dev);
+        oss << (name ? name : "") << '\t' << (desc ? desc : "") << '\n';
+    }
+
+    std::string devices = oss.str();
+    if (buffer != nullptr && buffer_size > 0) {
+        size_t copy_size = std::min(devices.size(), buffer_size - 1);
+        memcpy(buffer, devices.data(), copy_size);
+        buffer[copy_size] = '\0';
+    }
+    return devices.size();
+}

src/core/util.h

@@ -98,6 +98,9 @@ int sd_get_preview_interval();
 bool sd_should_preview_denoised();
 bool sd_should_preview_noisy();
 
+sd_graph_eval_callback_t sd_get_backend_eval_callback();
+void* sd_get_backend_eval_callback_data();
+
 // 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);
 

src/model.h

@@ -36,6 +36,7 @@ enum SDVersion {
     VERSION_WAN2_2_I2V,
     VERSION_WAN2_2_TI2V,
     VERSION_QWEN_IMAGE,
+    VERSION_QWEN_IMAGE_LAYERED,
     VERSION_ANIMA,
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
@@ -46,6 +47,7 @@ enum SDVersion {
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
     VERSION_LENS,
+    VERSION_MINIT2I,
     VERSION_LONGCAT,
     VERSION_PID,
     VERSION_IDEOGRAM4,
@@ -126,7 +128,7 @@ static inline bool sd_version_is_wan(SDVersion version) {
 }
 
 static inline bool sd_version_is_qwen_image(SDVersion version) {
-    if (version == VERSION_QWEN_IMAGE) {
+    if (version == VERSION_QWEN_IMAGE || version == VERSION_QWEN_IMAGE_LAYERED) {
         return true;
     }
     return false;
@@ -174,6 +176,13 @@ static inline bool sd_version_is_lens(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_minit2i(SDVersion version) {
+    if (version == VERSION_MINIT2I) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_pid(SDVersion version) {
     if (version == VERSION_PID) {
         return true;
@@ -247,6 +256,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_boogu_image(version) ||
         sd_version_is_ernie_image(version) ||
         sd_version_is_lens(version) ||
+        sd_version_is_minit2i(version) ||
         sd_version_is_longcat(version) ||
         sd_version_is_pid(version) ||
         sd_version_is_ideogram4(version) ||

src/model/common/rope.hpp

@@ -12,6 +12,16 @@ namespace Rope {
         ErnieImage,
     };
 
+    enum class RefIndexMode {
+        FIXED,
+        INCREASE,
+        DECREASE,
+    };
+
+    __STATIC_INLINE__ RefIndexMode ref_index_mode_from_bool(bool increase_ref_index) {
+        return increase_ref_index ? RefIndexMode::INCREASE : RefIndexMode::FIXED;
+    }
+
     template <class T>
     __STATIC_INLINE__ std::vector<T> linspace(T start, T end, int num) {
         std::vector<T> result(num);
@@ -346,7 +356,7 @@ namespace Rope {
                                                                    int axes_dim_num,
                                                                    int start_index,
                                                                    const std::vector<ggml_tensor*>& ref_latents,
-                                                                   bool increase_ref_index,
+                                                                   RefIndexMode ref_index_mode,
                                                                    float ref_index_scale,
                                                                    bool scale_rope,
                                                                    int base_offset = 0) {
@@ -357,13 +367,15 @@ namespace Rope {
         for (ggml_tensor* ref : ref_latents) {
             int h_offset = 0;
             int w_offset = 0;
-            if (!increase_ref_index) {
+            if (ref_index_mode == RefIndexMode::FIXED) {
                 if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) {
                     w_offset = curr_w_offset;
                 } else {
                     h_offset = curr_h_offset;
                 }
                 scale_rope = false;
+            } else if (ref_index_mode == RefIndexMode::DECREASE) {
+                index--;
             }
 
             auto ref_ids = gen_flux_img_ids(static_cast<int>(ref->ne[1]),
@@ -377,7 +389,7 @@ namespace Rope {
                                             scale_rope);
             ids          = concat_ids(ids, ref_ids, bs);
 
-            if (increase_ref_index) {
+            if (ref_index_mode == RefIndexMode::INCREASE) {
                 index++;
             }
 
@@ -395,7 +407,7 @@ namespace Rope {
                                                                    int context_len,
                                                                    std::set<int> txt_arange_dims,
                                                                    const std::vector<ggml_tensor*>& ref_latents,
-                                                                   bool increase_ref_index,
+                                                                   RefIndexMode ref_index_mode,
                                                                    float ref_index_scale,
                                                                    bool is_longcat) {
         int x_index = is_longcat ? 1 : 0;
@@ -406,7 +418,7 @@ namespace Rope {
 
         auto ids = concat_ids(txt_ids, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, increase_ref_index, ref_index_scale, false, offset);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, ref_index_mode, ref_index_scale, false, offset);
             ids           = concat_ids(ids, refs_ids, bs);
         }
         return ids;
@@ -420,7 +432,7 @@ namespace Rope {
                                                      int context_len,
                                                      std::set<int> txt_arange_dims,
                                                      const std::vector<ggml_tensor*>& ref_latents,
-                                                     bool increase_ref_index,
+                                                     RefIndexMode ref_index_mode,
                                                      float ref_index_scale,
                                                      int theta,
                                                      bool circular_h,
@@ -435,7 +447,7 @@ namespace Rope {
                                                            context_len,
                                                            txt_arange_dims,
                                                            ref_latents,
-                                                           increase_ref_index,
+                                                           ref_index_mode,
                                                            ref_index_scale,
                                                            is_longcat);
         std::vector<std::vector<int>> wrap_dims;
@@ -481,17 +493,64 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
+                                                                  int h,
+                                                                  int w,
+                                                                  int pt,
+                                                                  int ph,
+                                                                  int pw,
+                                                                  int bs,
+                                                                  int t_offset    = 0,
+                                                                  int h_offset    = 0,
+                                                                  int w_offset    = 0,
+                                                                  bool scale_rope = false) {
+        int t_len = (t + (pt / 2)) / pt;
+        int h_len = (h + (ph / 2)) / ph;
+        int w_len = (w + (pw / 2)) / pw;
+
+        std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0));
+
+        if (scale_rope) {
+            h_offset -= h_len / 2;
+            w_offset -= w_len / 2;
+        }
+
+        std::vector<float> t_ids = linspace<float>(1.f * t_offset, 1.f * t_len - 1 + t_offset, t_len);
+        std::vector<float> h_ids = linspace<float>(1.f * h_offset, 1.f * h_len - 1 + h_offset, h_len);
+        std::vector<float> w_ids = linspace<float>(1.f * w_offset, 1.f * w_len - 1 + w_offset, w_len);
+
+        for (int i = 0; i < t_len; ++i) {
+            for (int j = 0; j < h_len; ++j) {
+                for (int k = 0; k < w_len; ++k) {
+                    int idx         = i * h_len * w_len + j * w_len + k;
+                    vid_ids[idx][0] = t_ids[i];
+                    vid_ids[idx][1] = h_ids[j];
+                    vid_ids[idx][2] = w_ids[k];
+                }
+            }
+        }
+
+        std::vector<std::vector<float>> vid_ids_repeated(bs * vid_ids.size(), std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < vid_ids.size(); ++j) {
+                vid_ids_repeated[i * vid_ids.size() + j] = vid_ids[j];
+            }
+        }
+        return vid_ids_repeated;
+    }
+
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int t,
+                                                                         int h,
                                                                          int w,
                                                                          int patch_size,
                                                                          int bs,
                                                                          int context_len,
                                                                          const std::vector<ggml_tensor*>& ref_latents,
-                                                                         bool increase_ref_index) {
+                                                                         RefIndexMode ref_index_mode) {
         int h_len        = (h + (patch_size / 2)) / patch_size;
         int w_len        = (w + (patch_size / 2)) / patch_size;
-        int txt_id_start = std::max(h_len, w_len);
-        auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * context_len + txt_id_start, context_len);
+        int txt_id_start = std::max(h_len, w_len) / 2;
+        auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * txt_id_start + context_len - 1, context_len);
         std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3));
         for (int i = 0; i < bs; ++i) {
             for (int j = 0; j < txt_ids.size(); ++j) {
@@ -499,28 +558,30 @@ namespace Rope {
             }
         }
         int axes_dim_num = 3;
-        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, 0, 0, 0, true);
+        auto img_ids     = gen_vid_ids(t, h, w, 1, patch_size, patch_size, bs, 0, 0, 0, true);
         auto ids         = concat_ids(txt_ids_repeated, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, 1, ref_latents, increase_ref_index, 1.f, true);
-            ids           = concat_ids(ids, refs_ids, bs);
+            int ref_start_index = ref_index_mode == RefIndexMode::DECREASE ? 0 : 1;
+            auto refs_ids       = gen_refs_ids(patch_size, bs, axes_dim_num, ref_start_index, ref_latents, ref_index_mode, 1.f, true);
+            ids                 = concat_ids(ids, refs_ids, bs);
         }
         return ids;
     }
 
     // Generate qwen_image positional embeddings
-    __STATIC_INLINE__ std::vector<float> gen_qwen_image_pe(int h,
+    __STATIC_INLINE__ std::vector<float> gen_qwen_image_pe(int t,
+                                                           int h,
                                                            int w,
                                                            int patch_size,
                                                            int bs,
                                                            int context_len,
                                                            const std::vector<ggml_tensor*>& ref_latents,
-                                                           bool increase_ref_index,
+                                                           RefIndexMode ref_index_mode,
                                                            int theta,
                                                            bool circular_h,
                                                            bool circular_w,
                                                            const std::vector<int>& axes_dim) {
-        std::vector<std::vector<float>> ids = gen_qwen_image_ids(h, w, patch_size, bs, context_len, ref_latents, increase_ref_index);
+        std::vector<std::vector<float>> ids = gen_qwen_image_ids(t, h, w, patch_size, bs, context_len, ref_latents, ref_index_mode);
         std::vector<std::vector<int>> wrap_dims;
         // This logic simply stores the (pad and patch_adjusted) sizes of images so we can make sure rope correctly tiles
         if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
@@ -533,7 +594,7 @@ namespace Rope {
                 // Track per-token wrap lengths for the row/column axes so only spatial tokens become periodic.
                 wrap_dims.assign(axes_dim.size(), std::vector<int>(total_tokens / bs, 0));
                 size_t cursor           = context_len;  // ignore text tokens
-                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                const size_t img_tokens = static_cast<size_t>(t) * static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
                 for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
                     if (circular_h) {
                         wrap_dims[1][cursor + token_i] = h_len;
@@ -684,46 +745,6 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims, EmbedNDLayout::ErnieImage);
     }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
-                                                                  int h,
-                                                                  int w,
-                                                                  int pt,
-                                                                  int ph,
-                                                                  int pw,
-                                                                  int bs,
-                                                                  int t_offset = 0,
-                                                                  int h_offset = 0,
-                                                                  int w_offset = 0) {
-        int t_len = (t + (pt / 2)) / pt;
-        int h_len = (h + (ph / 2)) / ph;
-        int w_len = (w + (pw / 2)) / pw;
-
-        std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0));
-
-        std::vector<float> t_ids = linspace<float>(1.f * t_offset, 1.f * t_len - 1 + t_offset, t_len);
-        std::vector<float> h_ids = linspace<float>(1.f * h_offset, 1.f * h_len - 1 + h_offset, h_len);
-        std::vector<float> w_ids = linspace<float>(1.f * w_offset, 1.f * w_len - 1 + w_offset, w_len);
-
-        for (int i = 0; i < t_len; ++i) {
-            for (int j = 0; j < h_len; ++j) {
-                for (int k = 0; k < w_len; ++k) {
-                    int idx         = i * h_len * w_len + j * w_len + k;
-                    vid_ids[idx][0] = t_ids[i];
-                    vid_ids[idx][1] = h_ids[j];
-                    vid_ids[idx][2] = w_ids[k];
-                }
-            }
-        }
-
-        std::vector<std::vector<float>> vid_ids_repeated(bs * vid_ids.size(), std::vector<float>(3));
-        for (int i = 0; i < bs; ++i) {
-            for (int j = 0; j < vid_ids.size(); ++j) {
-                vid_ids_repeated[i * vid_ids.size() + j] = vid_ids[j];
-            }
-        }
-        return vid_ids_repeated;
-    }
-
     // Generate wan positional embeddings
     __STATIC_INLINE__ std::vector<float> gen_wan_pe(int t,
                                                     int h,
@@ -785,7 +806,8 @@ namespace Rope {
                                                                       int context_len,
                                                                       int seq_multi_of,
                                                                       const std::vector<ggml_tensor*>& ref_latents,
-                                                                      bool increase_ref_index) {
+                                                                      RefIndexMode ref_index_mode) {
+        SD_UNUSED(ref_index_mode);
         int padded_context_len = context_len + bound_mod(context_len, seq_multi_of);
         auto txt_ids           = std::vector<std::vector<float>>(bs * padded_context_len, std::vector<float>(3, 0.0f));
         for (int i = 0; i < bs * padded_context_len; i++) {
@@ -816,12 +838,12 @@ namespace Rope {
                                                         int context_len,
                                                         int seq_multi_of,
                                                         const std::vector<ggml_tensor*>& ref_latents,
-                                                        bool increase_ref_index,
+                                                        RefIndexMode ref_index_mode,
                                                         int theta,
                                                         bool circular_h,
                                                         bool circular_w,
                                                         const std::vector<int>& axes_dim) {
-        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, increase_ref_index);
+        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, ref_index_mode);
         std::vector<std::vector<int>> wrap_dims;
         if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
             int pad_h = (patch_size - (h % patch_size)) % patch_size;

src/model/diffusion/anima.hpp

@@ -612,7 +612,7 @@ namespace Anima {
                                           0,
                                           {},
                                           empty_ref_latents,
-                                          false,
+                                          Rope::RefIndexMode::FIXED,
                                           1.0f,
                                           false);
 

src/model/diffusion/control.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_DIFFUSION_CONTROL_HPP__
+#ifndef __SD_MODEL_DIFFUSION_CONTROL_HPP__
 #define __SD_MODEL_DIFFUSION_CONTROL_HPP__
 
 #include "model/common/block.hpp"

src/model/diffusion/dit.hpp

@@ -135,23 +135,23 @@ namespace DiT {
         return x;
     }
 
-    inline ggml_tensor* unpatchify(ggml_context* ctx,
-                                   ggml_tensor* x,
-                                   int64_t t_len,
-                                   int64_t h_len,
-                                   int64_t w_len,
-                                   int pt,
-                                   int ph,
-                                   int pw) {
-        // x: [N, t_len*h_len*w_len, pt*ph*pw*C]
+    inline ggml_tensor* unpatchify_3d(ggml_context* ctx,
+                                      ggml_tensor* x,
+                                      int64_t t_len,
+                                      int64_t h_len,
+                                      int64_t w_len,
+                                      int pt,
+                                      int ph,
+                                      int pw) {
+        // x: [N, t_len*h_len*w_len, C*pt*ph*pw]
         // return: [N*C, t_len*pt, h_len*ph, w_len*pw]
-        int64_t N = x->ne[3];
+        int64_t N = x->ne[2];
         int64_t C = x->ne[0] / pt / ph / pw;
 
         GGML_ASSERT(C * pt * ph * pw == x->ne[0]);
 
-        x = ggml_reshape_4d(ctx, x, C, pw * ph * pt, w_len * h_len * t_len, N);  // [N, t_len*h_len*w_len, pt*ph*pw, C]
-        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3));      // [N, C, t_len*h_len*w_len, pt*ph*pw]
+        x = ggml_reshape_4d(ctx, x, pw * ph * pt, C, w_len * h_len * t_len, N);  // [N, t_len*h_len*w_len, C, pt*ph*pw]
+        x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N, C, t_len*h_len*w_len, pt*ph*pw]
         x = ggml_reshape_4d(ctx, x, pw, ph * pt, w_len, h_len * t_len * C * N);  // [N*C*t_len*h_len, w_len, pt*ph, pw]
         x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));      // [N*C*t_len*h_len, pt*ph, w_len, pw]
         x = ggml_reshape_4d(ctx, x, pw * w_len, ph, pt, h_len * t_len * C * N);  // [N*C*t_len*h_len, pt, ph, w_len*pw]

src/model/diffusion/flux.hpp

@@ -1485,7 +1485,7 @@ namespace Flux {
                                  const sd::Tensor<float>& y_tensor                        = {},
                                  const sd::Tensor<float>& guidance_tensor                 = {},
                                  const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
-                                 bool increase_ref_index                                  = false,
+                                 Rope::RefIndexMode ref_index_mode                        = Rope::RefIndexMode::FIXED,
                                  std::vector<int> skip_layers                             = {},
                                  const sd::Tensor<float>& pulid_id_tensor                 = {},
                                  float pulid_id_weight                                    = 1.0f) {
@@ -1527,8 +1527,8 @@ namespace Flux {
             }
             std::set<int> txt_arange_dims;
             if (sd_version_is_flux2(version) || sd_version_is_sefi_image(version)) {
-                txt_arange_dims    = {3};
-                increase_ref_index = true;
+                txt_arange_dims = {3};
+                ref_index_mode  = Rope::RefIndexMode::INCREASE;
             } else if (version == VERSION_OVIS_IMAGE) {
                 txt_arange_dims = {1, 2};
             }
@@ -1539,7 +1539,7 @@ namespace Flux {
                                             static_cast<int>(context->ne[1]),
                                             txt_arange_dims,
                                             ref_latents,
-                                            increase_ref_index,
+                                            ref_index_mode,
                                             config.ref_index_scale,
                                             config.theta,
                                             circular_y_enabled,
@@ -1599,7 +1599,7 @@ namespace Flux {
                                   const sd::Tensor<float>& y                        = {},
                                   const sd::Tensor<float>& guidance                 = {},
                                   const std::vector<sd::Tensor<float>>& ref_latents = {},
-                                  bool increase_ref_index                           = false,
+                                  Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::FIXED,
                                   std::vector<int> skip_layers                      = std::vector<int>(),
                                   const sd::Tensor<float>& pulid_id                 = {},
                                   float pulid_id_weight                             = 1.0f) {
@@ -1610,7 +1610,7 @@ namespace Flux {
             // guidance: [N, ]
             // pulid_id: empty (no injection) or [N, num_id_tokens=32, kv_dim=2048]
             auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, increase_ref_index, skip_layers, pulid_id, pulid_id_weight);
+                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, ref_index_mode, skip_layers, pulid_id, pulid_id_weight);
             };
 
             auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
@@ -1632,7 +1632,7 @@ namespace Flux {
                            tensor_or_empty(diffusion_params.y),
                            tensor_or_empty(extra->guidance),
                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                           diffusion_params.increase_ref_index,
+                           diffusion_params.ref_index_mode,
                            extra->skip_layers ? *extra->skip_layers : empty_skip_layers,
                            tensor_or_empty(extra->pulid_id),
                            extra->pulid_id_weight);
@@ -1683,7 +1683,7 @@ namespace Flux {
                                        {},
                                        guidance,
                                        {},
-                                       false);
+                                       Rope::RefIndexMode::FIXED);
                 int64_t t1   = ggml_time_ms();
 
                 GGML_ASSERT(!out_opt.empty());

src/model/diffusion/hidream_o1.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
+#ifndef __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
 #define __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
 
 #include <algorithm>

src/model/diffusion/minit2i.hpp

@@ -0,0 +1,611 @@
+#ifndef __SD_MODEL_DIFFUSION_MINIT2I_HPP__
+#define __SD_MODEL_DIFFUSION_MINIT2I_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
+
+namespace MiniT2I {
+    constexpr int MINIT2I_GRAPH_SIZE = 196608;
+
+    struct MiniT2IConfig {
+        int64_t image_size         = 512;
+        int64_t patch_size         = 16;
+        int64_t in_channels        = 3;
+        int64_t txt_input_size     = 1024;
+        int64_t hidden_size        = 768;
+        int64_t txt_hidden_size    = 768;
+        int64_t cond_vec_size      = 768;
+        int64_t depth_double       = 17;
+        int64_t txt_preamble_depth = 2;
+        int64_t num_heads          = 12;
+        int64_t head_dim           = 64;
+        float mlp_ratio            = 2.6667f;
+        int64_t pca_channels       = 128;
+        int64_t prompt_length      = 256;
+        int64_t n_T                = 100;
+        float cfg_interval_start   = 0.0f;
+        float cfg_interval_end     = 1.0f;
+
+        static MiniT2IConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            MiniT2IConfig config;
+            config.depth_double       = 0;
+            config.txt_preamble_depth = 0;
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "img_embedder.proj1.weight") && tensor_storage.n_dims == 4) {
+                    config.patch_size   = tensor_storage.ne[0];
+                    config.in_channels  = tensor_storage.ne[2];
+                    config.pca_channels = tensor_storage.ne[3];
+                } else if (ends_with(name, "img_embedder.proj2.weight") && tensor_storage.n_dims == 4) {
+                    config.pca_channels = tensor_storage.ne[2];
+                    config.hidden_size  = tensor_storage.ne[3];
+                } else if (ends_with(name, "txt_embedder.weight") && tensor_storage.n_dims == 2) {
+                    config.txt_input_size  = tensor_storage.ne[0];
+                    config.txt_hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "pooled_embedder.weight") && tensor_storage.n_dims == 2) {
+                    config.cond_vec_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "double_blocks.0.img_qkv.weight") && tensor_storage.n_dims == 2) {
+                    int64_t inner3     = tensor_storage.ne[1];
+                    int64_t inner      = inner3 / 3;
+                    config.hidden_size = tensor_storage.ne[0];
+                    if (config.hidden_size == 768) {
+                        config.num_heads = 12;
+                        config.head_dim  = 64;
+                    } else if (config.hidden_size == 1248) {
+                        config.num_heads = 24;
+                        config.head_dim  = 52;
+                    } else if (inner > 0) {
+                        config.head_dim  = 64;
+                        config.num_heads = std::max<int64_t>(1, inner / config.head_dim);
+                    }
+                } else if (ends_with(name, "final_layer.linear.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area = config.patch_size * config.patch_size;
+                    config.hidden_size = tensor_storage.ne[0];
+                    config.in_channels = patch_area > 0 ? tensor_storage.ne[1] / patch_area : config.in_channels;
+                } else if (ends_with(name, "mask_token") && tensor_storage.n_dims >= 2) {
+                    config.prompt_length = tensor_storage.ne[1];
+                }
+
+                size_t pos = name.find("double_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int64_t idx         = atoi(items[1].c_str());
+                        config.depth_double = std::max<int64_t>(config.depth_double, idx + 1);
+                    }
+                }
+                pos = name.find("txt_preamble_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int64_t idx               = atoi(items[1].c_str());
+                        config.txt_preamble_depth = std::max<int64_t>(config.txt_preamble_depth, idx + 1);
+                    }
+                }
+            }
+
+            if (config.depth_double <= 0) {
+                config.depth_double = config.hidden_size == 1248 ? 23 : 17;
+            }
+            if (config.txt_preamble_depth <= 0) {
+                config.txt_preamble_depth = 2;
+            }
+            if (config.head_dim <= 0 || config.num_heads <= 0) {
+                config.head_dim  = config.hidden_size == 1248 ? 52 : 64;
+                config.num_heads = config.hidden_size / config.head_dim;
+            }
+            LOG_DEBUG("minit2i: hidden_size=%" PRId64 ", txt_hidden_size=%" PRId64 ", heads=%" PRId64 ", head_dim=%" PRId64 ", double_blocks=%" PRId64 ", txt_blocks=%" PRId64 ", patch=%" PRId64 ", in_channels=%" PRId64,
+                      config.hidden_size,
+                      config.txt_hidden_size,
+                      config.num_heads,
+                      config.head_dim,
+                      config.depth_double,
+                      config.txt_preamble_depth,
+                      config.patch_size,
+                      config.in_channels);
+            return config;
+        }
+    };
+
+    inline std::vector<float> make_2d_sincos_pos_embed(int grid_size, int dim) {
+        GGML_ASSERT(dim % 4 == 0);
+        int half_dim = dim / 2;
+        int quarter  = half_dim / 2;
+        std::vector<float> out(static_cast<size_t>(grid_size) * grid_size * dim);
+        std::vector<float> omega(quarter);
+        for (int i = 0; i < quarter; ++i) {
+            omega[i] = 1.0f / std::pow(10000.0f, static_cast<float>(i) / static_cast<float>(quarter));
+        }
+        for (int y = 0; y < grid_size; ++y) {
+            for (int x = 0; x < grid_size; ++x) {
+                size_t base = static_cast<size_t>(y * grid_size + x) * dim;
+                for (int i = 0; i < quarter; ++i) {
+                    float ay                           = y * omega[i];
+                    float ax                           = x * omega[i];
+                    out[base + i]                      = std::sin(ax);
+                    out[base + quarter + i]            = std::cos(ax);
+                    out[base + half_dim + i]           = std::sin(ay);
+                    out[base + half_dim + quarter + i] = std::cos(ay);
+                }
+            }
+        }
+        return out;
+    }
+
+    inline std::vector<float> make_text_rope(int length, int head_dim) {
+        return Rope::flatten(Rope::rope(Rope::linspace(0.f, static_cast<float>(length - 1), length), head_dim, 10000.f));
+    }
+
+    inline std::vector<float> make_vision_rope(int side, int head_dim) {
+        GGML_ASSERT(head_dim % 4 == 0);
+        int dim     = head_dim / 2;
+        int quarter = dim / 2;
+        int length  = side * side;
+        std::vector<float> out(static_cast<size_t>(length) * (head_dim / 2) * 4);
+        std::vector<float> freqs(quarter);
+        for (int i = 0; i < quarter; ++i) {
+            freqs[i] = 1.0f / std::pow(10000.0f, static_cast<float>(2 * i) / static_cast<float>(dim));
+        }
+        for (int y = 0; y < side; ++y) {
+            for (int x = 0; x < side; ++x) {
+                int pos     = y * side + x;
+                size_t base = static_cast<size_t>(pos) * (head_dim / 2) * 4;
+                for (int i = 0; i < quarter; ++i) {
+                    float ay        = y * freqs[i];
+                    float ax        = x * freqs[i];
+                    float angles[2] = {ay, ax};
+                    for (int axis = 0; axis < 2; ++axis) {
+                        int j                 = axis * quarter + i;
+                        out[base + 4 * j]     = std::cos(angles[axis]);
+                        out[base + 4 * j + 1] = -std::sin(angles[axis]);
+                        out[base + 4 * j + 2] = std::sin(angles[axis]);
+                        out[base + 4 * j + 3] = std::cos(angles[axis]);
+                    }
+                }
+            }
+        }
+        return out;
+    }
+
+    struct SwiGLUMlp : public GGMLBlock {
+        SwiGLUMlp(int64_t in_features, int64_t hidden_features) {
+            int64_t hidden_dim = ((hidden_features + 7) / 8) * 8;
+            blocks["w1"]       = std::make_shared<Linear>(in_features, hidden_dim, false);
+            blocks["w3"]       = std::make_shared<Linear>(in_features, hidden_dim, false);
+            blocks["w2"]       = std::make_shared<Linear>(hidden_dim, in_features, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1   = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w3   = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+            auto w2   = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto gate = ggml_silu(ctx->ggml_ctx, w1->forward(ctx, x));
+            auto up   = w3->forward(ctx, x);
+            return w2->forward(ctx, ggml_mul(ctx->ggml_ctx, gate, up));
+        }
+    };
+
+    struct BottleneckPatchEmbed : public GGMLBlock {
+        int64_t patch_size;
+
+        BottleneckPatchEmbed(int64_t patch_size, int64_t in_channels, int64_t pca_channels, int64_t hidden_size)
+            : patch_size(patch_size) {
+            blocks["proj1"] = std::make_shared<Conv2d>(in_channels,
+                                                       pca_channels,
+                                                       std::pair<int, int>{static_cast<int>(patch_size), static_cast<int>(patch_size)},
+                                                       std::pair<int, int>{static_cast<int>(patch_size), static_cast<int>(patch_size)},
+                                                       std::pair<int, int>{0, 0},
+                                                       std::pair<int, int>{1, 1},
+                                                       false);
+            blocks["proj2"] = std::make_shared<Conv2d>(pca_channels,
+                                                       hidden_size,
+                                                       std::pair<int, int>{1, 1},
+                                                       std::pair<int, int>{1, 1},
+                                                       std::pair<int, int>{0, 0},
+                                                       std::pair<int, int>{1, 1},
+                                                       true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj1 = std::dynamic_pointer_cast<Conv2d>(blocks["proj1"]);
+            auto proj2 = std::dynamic_pointer_cast<Conv2d>(blocks["proj2"]);
+            x          = proj1->forward(ctx, x);
+            x          = proj2->forward(ctx, x);
+            x          = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0] * x->ne[1], x->ne[2], x->ne[3]);
+            x          = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
+            return x;
+        }
+    };
+
+    struct TimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size;
+
+        TimestepEmbedder(int64_t hidden_size, int frequency_embedding_size = 256)
+            : frequency_embedding_size(frequency_embedding_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true, 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 t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10000, 1.0f);
+            t_emb      = mlp_0->forward(ctx, t_emb);
+            t_emb      = ggml_silu_inplace(ctx->ggml_ctx, t_emb);
+            return mlp_2->forward(ctx, t_emb);
+        }
+    };
+
+    inline std::vector<ggml_tensor*> split_qkv(ggml_context* ctx, ggml_tensor* qkv, int64_t num_heads, int64_t head_dim) {
+        int64_t N = qkv->ne[2];
+        int64_t L = qkv->ne[1];
+        auto q    = ggml_view_4d(ctx, qkv, head_dim, num_heads, L, N,
+                                 qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], 0);
+        auto k    = ggml_view_4d(ctx, qkv, head_dim, num_heads, L, N,
+                                 qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], qkv->nb[0] * head_dim * num_heads);
+        auto v    = ggml_view_4d(ctx, qkv, head_dim, num_heads, L, N,
+                                 qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], qkv->nb[0] * head_dim * num_heads * 2);
+        return {q, k, v};
+    }
+
+    struct PlainTextTransformerBlock : public GGMLBlock {
+        int64_t num_heads;
+        int64_t head_dim;
+
+        PlainTextTransformerBlock(int64_t hidden_size, int64_t num_heads, int64_t head_dim, float mlp_ratio)
+            : num_heads(num_heads), head_dim(head_dim) {
+            int64_t inner_dim   = num_heads * head_dim;
+            blocks["norm1"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["norm2"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["qkv"]       = std::make_shared<Linear>(hidden_size, inner_dim * 3, true);
+            blocks["attn_proj"] = std::make_shared<Linear>(inner_dim, hidden_size, true);
+            blocks["mlp"]       = std::make_shared<SwiGLUMlp>(hidden_size, static_cast<int64_t>(hidden_size * mlp_ratio));
+            blocks["q_norm"]    = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"]    = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* txt, ggml_tensor* pe) {
+            auto norm1     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+            auto norm2     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto qkv_proj  = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto attn_proj = std::dynamic_pointer_cast<Linear>(blocks["attn_proj"]);
+            auto mlp       = std::dynamic_pointer_cast<SwiGLUMlp>(blocks["mlp"]);
+            auto q_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+
+            auto qkv = split_qkv(ctx->ggml_ctx, qkv_proj->forward(ctx, norm1->forward(ctx, txt)), num_heads, head_dim);
+            auto q   = q_norm->forward(ctx, qkv[0]);
+            auto k   = k_norm->forward(ctx, qkv[1]);
+            auto v   = qkv[2];
+            auto out = Rope::attention(ctx, q, k, v, pe, nullptr, 1.0f, false);
+            txt      = ggml_add(ctx->ggml_ctx, txt, attn_proj->forward(ctx, out));
+            txt      = ggml_add(ctx->ggml_ctx, txt, mlp->forward(ctx, norm2->forward(ctx, txt)));
+            return txt;
+        }
+    };
+
+    struct DoubleStreamDiTBlock : public GGMLBlock {
+        int64_t num_heads;
+        int64_t head_dim;
+
+        DoubleStreamDiTBlock(int64_t hidden_size, int64_t txt_hidden_size, int64_t num_heads, int64_t head_dim, float mlp_ratio)
+            : num_heads(num_heads), head_dim(head_dim) {
+            int64_t inner_dim       = num_heads * head_dim;
+            blocks["img_norm1"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["img_norm2"]     = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["txt_norm1"]     = std::make_shared<RMSNorm>(txt_hidden_size, 1e-6f);
+            blocks["txt_norm2"]     = std::make_shared<RMSNorm>(txt_hidden_size, 1e-6f);
+            blocks["img_qkv"]       = std::make_shared<Linear>(hidden_size, inner_dim * 3, true);
+            blocks["txt_qkv"]       = std::make_shared<Linear>(txt_hidden_size, inner_dim * 3, true);
+            blocks["q_norm"]        = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"]        = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["img_attn_proj"] = std::make_shared<Linear>(inner_dim, hidden_size, true);
+            blocks["txt_attn_proj"] = std::make_shared<Linear>(inner_dim, txt_hidden_size, true);
+            blocks["img_mlp"]       = std::make_shared<SwiGLUMlp>(hidden_size, static_cast<int64_t>(hidden_size * mlp_ratio));
+            blocks["txt_mlp"]       = std::make_shared<SwiGLUMlp>(txt_hidden_size, static_cast<int64_t>(txt_hidden_size * mlp_ratio));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* pe) {
+            auto img_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm2"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm2"]);
+            auto img_qkv_p = std::dynamic_pointer_cast<Linear>(blocks["img_qkv"]);
+            auto txt_qkv_p = std::dynamic_pointer_cast<Linear>(blocks["txt_qkv"]);
+            auto q_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto img_proj  = std::dynamic_pointer_cast<Linear>(blocks["img_attn_proj"]);
+            auto txt_proj  = std::dynamic_pointer_cast<Linear>(blocks["txt_attn_proj"]);
+            auto img_mlp   = std::dynamic_pointer_cast<SwiGLUMlp>(blocks["img_mlp"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<SwiGLUMlp>(blocks["txt_mlp"]);
+
+            int64_t li = img->ne[1];
+            int64_t lt = txt->ne[1];
+
+            auto img_qkv = split_qkv(ctx->ggml_ctx, img_qkv_p->forward(ctx, img_norm1->forward(ctx, img)), num_heads, head_dim);
+            auto txt_qkv = split_qkv(ctx->ggml_ctx, txt_qkv_p->forward(ctx, txt_norm1->forward(ctx, txt)), num_heads, head_dim);
+
+            auto q = ggml_concat(ctx->ggml_ctx, q_norm->forward(ctx, txt_qkv[0]), q_norm->forward(ctx, img_qkv[0]), 2);
+            auto k = ggml_concat(ctx->ggml_ctx, k_norm->forward(ctx, txt_qkv[1]), k_norm->forward(ctx, img_qkv[1]), 2);
+            auto v = ggml_concat(ctx->ggml_ctx, txt_qkv[2], img_qkv[2], 2);
+
+            auto out     = Rope::attention(ctx, q, k, v, pe, nullptr, 1.0f, false);
+            auto out_txt = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, lt);
+            auto out_img = ggml_ext_slice(ctx->ggml_ctx, out, 1, lt, lt + li);
+
+            img = ggml_add(ctx->ggml_ctx, img, img_proj->forward(ctx, out_img));
+            txt = ggml_add(ctx->ggml_ctx, txt, txt_proj->forward(ctx, out_txt));
+            img = ggml_add(ctx->ggml_ctx, img, img_mlp->forward(ctx, img_norm2->forward(ctx, img)));
+            txt = ggml_add(ctx->ggml_ctx, txt, txt_mlp->forward(ctx, txt_norm2->forward(ctx, txt)));
+            return {img, txt};
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t patch_size, int64_t out_channels) {
+            blocks["norm_final"] = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["linear"]     = std::make_shared<Linear>(hidden_size, patch_size * patch_size * out_channels, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto norm_final = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_final"]);
+            auto linear     = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, norm_final->forward(ctx, x));
+        }
+    };
+
+    struct MMJiT : public GGMLBlock {
+        MiniT2IConfig config;
+
+        MMJiT(const MiniT2IConfig& config)
+            : config(config) {
+            blocks["img_embedder"]    = std::make_shared<BottleneckPatchEmbed>(config.patch_size, config.in_channels, config.pca_channels, config.hidden_size);
+            blocks["txt_embedder"]    = std::make_shared<Linear>(config.txt_input_size, config.txt_hidden_size, false);
+            blocks["t_embedder"]      = std::make_shared<TimestepEmbedder>(config.cond_vec_size);
+            blocks["pooled_embedder"] = std::make_shared<Linear>(config.txt_input_size, config.cond_vec_size, false);
+            for (int64_t i = 0; i < config.txt_preamble_depth; ++i) {
+                blocks["txt_preamble_blocks." + std::to_string(i)] = std::make_shared<PlainTextTransformerBlock>(config.txt_hidden_size, config.num_heads, config.head_dim, config.mlp_ratio);
+            }
+            for (int64_t i = 0; i < config.depth_double; ++i) {
+                blocks["double_blocks." + std::to_string(i)] = std::make_shared<DoubleStreamDiTBlock>(config.hidden_size, config.txt_hidden_size, config.num_heads, config.head_dim, config.mlp_ratio);
+            }
+            blocks["final_layer"] = std::make_shared<FinalLayer>(config.hidden_size, config.patch_size, config.in_channels);
+        }
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGMLBlock::init_params(ctx, tensor_storage_map, prefix);
+            enum ggml_type wtype = get_type(prefix + "mask_token", tensor_storage_map, GGML_TYPE_F32);
+            params["mask_token"] = ggml_new_tensor_3d(ctx, wtype, config.txt_input_size, 1, 1);
+        }
+
+        ggml_tensor* apply_text_mask(GGMLRunnerContext* ctx, ggml_tensor* context, ggml_tensor* mask) {
+            if (mask == nullptr) {
+                return context;
+            }
+            mask            = ggml_reshape_3d(ctx->ggml_ctx, mask, 1, mask->ne[0], mask->ne[1]);
+            mask            = ggml_repeat(ctx->ggml_ctx, mask, context);
+            auto keep       = ggml_mul(ctx->ggml_ctx, context, mask);
+            auto inv        = ggml_sub(ctx->ggml_ctx, ggml_ext_ones_like(ctx->ggml_ctx, mask), mask);
+            auto mask_token = ggml_repeat(ctx->ggml_ctx, params["mask_token"], context);
+            return ggml_add(ctx->ggml_ctx, keep, ggml_mul(ctx->ggml_ctx, mask_token, inv));
+        }
+
+        ggml_tensor* pool_context(GGMLRunnerContext* ctx, ggml_tensor* context) {
+            int64_t dim = context->ne[0];
+            int64_t len = context->ne[1];
+            int64_t N   = context->ne[2];
+            auto x      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, context, 1, 0, 2, 3));
+            x           = ggml_reshape_3d(ctx->ggml_ctx, x, len, dim, N);
+            x           = ggml_mean(ctx->ggml_ctx, x);
+            x           = ggml_reshape_2d(ctx->ggml_ctx, x, dim, N);
+            return x;
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* img,
+                             ggml_tensor* context,
+                             ggml_tensor* mask,
+                             ggml_tensor* pos_embed,
+                             ggml_tensor* txt_pe,
+                             ggml_tensor* joint_pe) {
+            auto img_embedder = std::dynamic_pointer_cast<BottleneckPatchEmbed>(blocks["img_embedder"]);
+            auto txt_embedder = std::dynamic_pointer_cast<Linear>(blocks["txt_embedder"]);
+            auto final_layer  = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
+
+            int64_t W  = img->ne[0];
+            int64_t H  = img->ne[1];
+            int64_t hp = H / config.patch_size;
+            int64_t wp = W / config.patch_size;
+
+            context = apply_text_mask(ctx, context, mask);
+            auto x  = img_embedder->forward(ctx, img);
+            x       = ggml_add(ctx->ggml_ctx, x, pos_embed);
+
+            auto txt = txt_embedder->forward(ctx, context);
+            for (int64_t i = 0; i < config.txt_preamble_depth; ++i) {
+                auto block = std::dynamic_pointer_cast<PlainTextTransformerBlock>(blocks["txt_preamble_blocks." + std::to_string(i)]);
+                txt        = block->forward(ctx, txt, txt_pe);
+                sd::ggml_graph_cut::mark_graph_cut(txt, "minit2i.txt_preamble_blocks." + std::to_string(i), "txt");
+            }
+            for (int64_t i = 0; i < config.depth_double; ++i) {
+                auto block = std::dynamic_pointer_cast<DoubleStreamDiTBlock>(blocks["double_blocks." + std::to_string(i)]);
+                auto out   = block->forward(ctx, x, txt, joint_pe);
+                x          = out.first;
+                txt        = out.second;
+                sd::ggml_graph_cut::mark_graph_cut(x, "minit2i.double_blocks." + std::to_string(i), "x");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "minit2i.double_blocks." + std::to_string(i), "txt");
+            }
+            auto combined = ggml_concat(ctx->ggml_ctx, txt, x, 1);
+            auto out      = final_layer->forward(ctx, combined);
+            auto img_out  = ggml_ext_slice(ctx->ggml_ctx, out, 1, txt->ne[1], txt->ne[1] + x->ne[1]);
+            return DiT::unpatchify(ctx->ggml_ctx, img_out, hp, wp, static_cast<int>(config.patch_size), static_cast<int>(config.patch_size), false);
+        }
+    };
+
+    struct MiniT2IRunner : public DiffusionModelRunner {
+        MiniT2IConfig config;
+        MMJiT model;
+        ggml_context* position_cache_ctx            = nullptr;
+        ggml_backend_buffer_t position_cache_buffer = nullptr;
+        ggml_tensor* cached_pos_embed               = nullptr;
+        ggml_tensor* cached_txt_pe                  = nullptr;
+        ggml_tensor* cached_joint_pe                = nullptr;
+        int64_t cached_img_side                     = -1;
+        int64_t cached_txt_len                      = -1;
+        int64_t cached_hidden_size                  = -1;
+        int64_t cached_head_dim                     = -1;
+
+        MiniT2IRunner(ggml_backend_t backend,
+                      const String2TensorStorage& tensor_storage_map      = {},
+                      const std::string prefix                            = "",
+                      std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(MiniT2IConfig::detect_from_weights(tensor_storage_map, this->prefix)),
+              model(config) {
+            model.init(params_ctx, tensor_storage_map, this->prefix);
+        }
+
+        ~MiniT2IRunner() override {
+            free_position_cache();
+        }
+
+        std::string get_desc() override {
+            return "MiniT2I";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        void free_position_cache() {
+            if (position_cache_buffer != nullptr) {
+                ggml_backend_buffer_free(position_cache_buffer);
+                position_cache_buffer = nullptr;
+            }
+            if (position_cache_ctx != nullptr) {
+                ggml_free(position_cache_ctx);
+                position_cache_ctx = nullptr;
+            }
+            cached_pos_embed   = nullptr;
+            cached_txt_pe      = nullptr;
+            cached_joint_pe    = nullptr;
+            cached_img_side    = -1;
+            cached_txt_len     = -1;
+            cached_hidden_size = -1;
+            cached_head_dim    = -1;
+        }
+
+        void ensure_position_cache(int64_t img_side, int64_t txt_len) {
+            if (cached_img_side == img_side &&
+                cached_txt_len == txt_len &&
+                cached_hidden_size == config.hidden_size &&
+                cached_head_dim == config.head_dim &&
+                cached_pos_embed != nullptr &&
+                cached_txt_pe != nullptr &&
+                cached_joint_pe != nullptr) {
+                return;
+            }
+
+            free_position_cache();
+
+            auto pos_embed_vec = make_2d_sincos_pos_embed(static_cast<int>(img_side), static_cast<int>(config.hidden_size));
+            auto txt_pe_vec    = make_text_rope(static_cast<int>(txt_len), static_cast<int>(config.head_dim));
+            auto img_pe_vec    = make_vision_rope(static_cast<int>(img_side), static_cast<int>(config.head_dim));
+            auto joint_pe_vec  = txt_pe_vec;
+            joint_pe_vec.insert(joint_pe_vec.end(), img_pe_vec.begin(), img_pe_vec.end());
+
+            ggml_init_params params;
+            params.mem_size    = static_cast<size_t>(3 * ggml_tensor_overhead());
+            params.mem_buffer  = nullptr;
+            params.no_alloc    = true;
+            position_cache_ctx = ggml_init(params);
+            GGML_ASSERT(position_cache_ctx != nullptr);
+
+            cached_pos_embed = ggml_new_tensor_3d(position_cache_ctx, GGML_TYPE_F32, config.hidden_size, img_side * img_side, 1);
+            ggml_set_name(cached_pos_embed, "minit2i.pos_embed");
+            cached_txt_pe = ggml_new_tensor_4d(position_cache_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, txt_len);
+            ggml_set_name(cached_txt_pe, "minit2i.txt_pe");
+            cached_joint_pe = ggml_new_tensor_4d(position_cache_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, txt_len + img_side * img_side);
+            ggml_set_name(cached_joint_pe, "minit2i.joint_pe");
+
+            position_cache_buffer = ggml_backend_alloc_ctx_tensors(position_cache_ctx, runtime_backend);
+            GGML_ASSERT(position_cache_buffer != nullptr);
+            ggml_backend_buffer_set_usage(position_cache_buffer, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+            ggml_backend_tensor_set(cached_pos_embed, pos_embed_vec.data(), 0, ggml_nbytes(cached_pos_embed));
+            ggml_backend_tensor_set(cached_txt_pe, txt_pe_vec.data(), 0, ggml_nbytes(cached_txt_pe));
+            ggml_backend_tensor_set(cached_joint_pe, joint_pe_vec.data(), 0, ggml_nbytes(cached_joint_pe));
+            ggml_backend_synchronize(runtime_backend);
+
+            cached_img_side    = img_side;
+            cached_txt_len     = txt_len;
+            cached_hidden_size = config.hidden_size;
+            cached_head_dim    = config.head_dim;
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const sd::Tensor<float>& mask_tensor) {
+            ggml_cgraph* gf      = new_graph_custom(MINIT2I_GRAPH_SIZE);
+            ggml_tensor* x       = make_input(x_tensor);
+            ggml_tensor* context = make_input(context_tensor);
+            ggml_tensor* mask    = make_input(mask_tensor);
+            SD_UNUSED(timesteps_tensor);
+
+            int64_t W        = x->ne[0];
+            int64_t H        = x->ne[1];
+            int64_t img_side = H / config.patch_size;
+            int64_t txt_len  = context->ne[1];
+            ensure_position_cache(img_side, txt_len);
+
+            auto runner_ctx = get_context();
+            auto out        = model.forward(&runner_ctx, x, context, mask, cached_pos_embed, cached_txt_pe, cached_joint_pe);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const sd::Tensor<float>& mask) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, mask);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            GGML_ASSERT(diffusion_params.context != nullptr);
+            const auto* extra = diffusion_extra_as<MiniT2IDiffusionExtra>(diffusion_params);
+            GGML_ASSERT(extra->mask != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *diffusion_params.context,
+                           *extra->mask);
+        }
+    };
+}  // namespace MiniT2I
+
+#endif  // __SD_MODEL_DIFFUSION_MINIT2I_HPP__

src/model/diffusion/model.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_DIFFUSION_MODEL_HPP__
+#ifndef __SD_MODEL_DIFFUSION_MODEL_HPP__
 #define __SD_MODEL_DIFFUSION_MODEL_HPP__
 
 #include <string>
@@ -7,6 +7,7 @@
 
 #include "core/ggml_extend.hpp"
 #include "core/tensor_ggml.hpp"
+#include "model/common/rope.hpp"
 #include "model_manager.h"
 
 struct UNetDiffusionExtra {
@@ -52,6 +53,10 @@ struct LTXAVDiffusionExtra {
     const sd::Tensor<float>* video_positions = nullptr;
 };
 
+struct MiniT2IDiffusionExtra {
+    const sd::Tensor<float>* mask = nullptr;
+};
+
 using DiffusionExtraParams = std::variant<std::monostate,
                                           UNetDiffusionExtra,
                                           SkipLayerDiffusionExtra,
@@ -59,7 +64,8 @@ using DiffusionExtraParams = std::variant<std::monostate,
                                           AnimaDiffusionExtra,
                                           WanDiffusionExtra,
                                           HiDreamO1DiffusionExtra,
-                                          LTXAVDiffusionExtra>;
+                                          LTXAVDiffusionExtra,
+                                          MiniT2IDiffusionExtra>;
 
 struct DiffusionParams {
     const sd::Tensor<float>* x                        = nullptr;
@@ -68,7 +74,7 @@ struct DiffusionParams {
     const sd::Tensor<float>* c_concat                 = nullptr;
     const sd::Tensor<float>* y                        = nullptr;
     const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
-    bool increase_ref_index                           = false;
+    Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::FIXED;
     DiffusionExtraParams extra                        = std::monostate{};
 };
 

src/model/diffusion/qwen_image.hpp

@@ -4,6 +4,7 @@
 #include <memory>
 
 #include "model/common/block.hpp"
+#include "model/diffusion/dit.hpp"
 #include "model/diffusion/flux.hpp"
 #include "model/diffusion/model.hpp"
 #include "model_loader.h"
@@ -23,6 +24,7 @@ namespace Qwen {
         std::vector<int> axes_dim   = {16, 56, 56};
         int axes_dim_sum            = 128;
         bool zero_cond_t            = false;
+        bool use_additional_t_cond  = false;
 
         static QwenImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
             QwenImageConfig config;
@@ -88,19 +90,33 @@ namespace Qwen {
     };
 
     struct QwenTimestepProjEmbeddings : public GGMLBlock {
+    protected:
+        bool use_additional_t_cond = false;
+
     public:
-        QwenTimestepProjEmbeddings(int64_t embedding_dim) {
+        QwenTimestepProjEmbeddings(int64_t embedding_dim, bool use_additional_t_cond = false)
+            : use_additional_t_cond(use_additional_t_cond) {
             blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
+            if (use_additional_t_cond) {
+                blocks["addition_t_embedding"] = std::shared_ptr<GGMLBlock>(new Embedding(2, embedding_dim));
+            }
         }
 
         ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* timesteps) {
+                             ggml_tensor* timesteps,
+                             ggml_tensor* addition_t_cond = nullptr) {
             // timesteps: [N,]
             // return: [N, embedding_dim]
             auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
 
             auto timesteps_proj = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 256, 10000, 1.f);
             auto timesteps_emb  = timestep_embedder->forward(ctx, timesteps_proj);
+            if (use_additional_t_cond) {
+                GGML_ASSERT(addition_t_cond != nullptr);
+                auto addition_t_embedding = std::dynamic_pointer_cast<Embedding>(blocks["addition_t_embedding"]);
+                auto addition_t_emb       = addition_t_embedding->forward(ctx, addition_t_cond);
+                timesteps_emb             = ggml_add(ctx->ggml_ctx, timesteps_emb, addition_t_emb);
+            }
             return timesteps_emb;
         }
     };
@@ -402,7 +418,7 @@ namespace Qwen {
         QwenImageModel(QwenImageConfig config)
             : config(config) {
             int64_t inner_dim         = config.num_attention_heads * config.attention_head_dim;
-            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim));
+            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim, config.use_additional_t_cond));
             blocks["txt_norm"]        = std::shared_ptr<GGMLBlock>(new RMSNorm(config.joint_attention_dim, 1e-6f));
             blocks["img_in"]          = std::shared_ptr<GGMLBlock>(new Linear(config.in_channels, inner_dim));
             blocks["txt_in"]          = std::shared_ptr<GGMLBlock>(new Linear(config.joint_attention_dim, inner_dim));
@@ -424,6 +440,7 @@ namespace Qwen {
         ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
                                   ggml_tensor* x,
                                   ggml_tensor* timestep,
+                                  ggml_tensor* addition_t_cond,
                                   ggml_tensor* context,
                                   ggml_tensor* pe,
                                   ggml_tensor* modulate_index = nullptr) {
@@ -434,9 +451,9 @@ namespace Qwen {
             auto norm_out        = std::dynamic_pointer_cast<AdaLayerNormContinuous>(blocks["norm_out"]);
             auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
 
-            auto t_emb = time_text_embed->forward(ctx, timestep);
+            auto t_emb = time_text_embed->forward(ctx, timestep, addition_t_cond);
             if (config.zero_cond_t) {
-                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep), addition_t_cond);
                 t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
             }
             auto img = img_in->forward(ctx, x);
@@ -469,33 +486,50 @@ namespace Qwen {
         ggml_tensor* forward(GGMLRunnerContext* ctx,
                              ggml_tensor* x,
                              ggml_tensor* timestep,
+                             ggml_tensor* addition_t_cond,
                              ggml_tensor* context,
                              ggml_tensor* pe,
                              std::vector<ggml_tensor*> ref_latents = {},
                              ggml_tensor* modulate_index           = nullptr) {
             // Forward pass of DiT.
-            // x: [N, C, H, W]
+            // x: [N, C, H, W] or [N*C, T, H, W]
             // timestep: [N,]
             // context: [N, L, D]
             // pe: [L, d_head/2, 2, 2]
-            // return: [N, C, H, W]
+            // return: [N, C, H, W] or [N*C, T, H, W]
 
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-            int64_t C = x->ne[2];
-            int64_t N = x->ne[3];
+            int64_t W          = x->ne[0];
+            int64_t H          = x->ne[1];
+            int64_t T          = 1;
+            int64_t N          = addition_t_cond != nullptr ? addition_t_cond->ne[0] : x->ne[3];
+            bool has_time_axis = false;
+            if (x->ne[3] != 1) {
+                T             = x->ne[2];
+                has_time_axis = true;
+            }
 
-            auto img           = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size);
+            auto patchify_input = [&](ggml_tensor* input) -> ggml_tensor* {
+                input = DiT::pad_to_patch_size(ctx, input, config.patch_size, config.patch_size);
+                if (!has_time_axis) {
+                    return DiT::patchify(ctx->ggml_ctx, input, config.patch_size, config.patch_size);
+                }
+                if (input->ne[3] == 1) {
+                    input = ggml_reshape_4d(ctx->ggml_ctx, input, input->ne[0], input->ne[1], 1, input->ne[2]);
+                }
+                return DiT::patchify(ctx->ggml_ctx, input, 1, config.patch_size, config.patch_size, N);
+            };
+
+            auto img           = patchify_input(x);
             int64_t img_tokens = img->ne[1];
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = DiT::pad_and_patchify(ctx, ref, config.patch_size, config.patch_size);
+                    ref = patchify_input(ref);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
 
-            auto out = forward_orig(ctx, img, timestep, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]
+            auto out = forward_orig(ctx, img, timestep, addition_t_cond, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]
 
             if (out->ne[1] > img_tokens) {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [num_tokens, N, C * patch_size * patch_size]
@@ -503,7 +537,17 @@ namespace Qwen {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
             }
 
-            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, config.patch_size, config.patch_size);  // [N, C, H, W]
+            if (has_time_axis) {
+                int pad_h = (config.patch_size - H % config.patch_size) % config.patch_size;
+                int pad_w = (config.patch_size - W % config.patch_size) % config.patch_size;
+                int h_len = static_cast<int>((H + pad_h) / config.patch_size);
+                int w_len = static_cast<int>((W + pad_w) / config.patch_size);
+                out       = DiT::unpatchify_3d(ctx->ggml_ctx, out, T, h_len, w_len, 1, config.patch_size, config.patch_size);
+                out       = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H);  // [N*C, T, H, W + pad_w]
+                out       = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W);  // [N*C, T, H, W]
+            } else {
+                out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, config.patch_size, config.patch_size);  // [N, C, H, W]
+            }
 
             return out;
         }
@@ -515,6 +559,7 @@ namespace Qwen {
         QwenImageModel qwen_image;
         std::vector<float> pe_vec;
         std::vector<float> modulate_index_vec;
+        std::vector<int32_t> additional_t_cond_vec;
         SDVersion version;
 
         QwenImageRunner(ggml_backend_t backend,
@@ -524,9 +569,13 @@ namespace Qwen {
                         bool zero_cond_t                                    = false,
                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
             : DiffusionModelRunner(backend, prefix, weight_manager),
-              config(QwenImageConfig::detect_from_weights(tensor_storage_map, prefix)) {
+              config(QwenImageConfig::detect_from_weights(tensor_storage_map, prefix)),
+              version(version) {
             config.zero_cond_t = config.zero_cond_t || zero_cond_t;
-            qwen_image         = QwenImageModel(config);
+            if (version == VERSION_QWEN_IMAGE_LAYERED) {
+                config.use_additional_t_cond = true;
+            }
+            qwen_image = QwenImageModel(config);
             qwen_image.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -542,11 +591,11 @@ namespace Qwen {
                                  const sd::Tensor<float>& timesteps_tensor,
                                  const sd::Tensor<float>& context_tensor,
                                  const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
-                                 bool increase_ref_index                                  = false) {
+                                 Rope::RefIndexMode ref_index_mode                        = Rope::RefIndexMode::INCREASE) {
             ggml_cgraph* gf        = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE);
             ggml_tensor* x         = make_input(x_tensor);
             ggml_tensor* timesteps = make_input(timesteps_tensor);
-            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(x->ne[3] == 1 || x_tensor.dim() == 5);
             GGML_ASSERT(!context_tensor.empty());
             ggml_tensor* context = make_input(context_tensor);
             std::vector<ggml_tensor*> ref_latents;
@@ -555,13 +604,29 @@ namespace Qwen {
                 ref_latents.push_back(make_input(ref_latent_tensor));
             }
 
-            pe_vec      = Rope::gen_qwen_image_pe(static_cast<int>(x->ne[1]),
+            int batch_size = static_cast<int>(x->ne[3]);
+            int time_len   = 1;
+            if (x_tensor.dim() == 5) {
+                time_len   = static_cast<int>(x_tensor.shape()[2]);
+                batch_size = static_cast<int>(x_tensor.shape()[4]);
+            }
+
+            ggml_tensor* addition_t_cond = nullptr;
+            if (version == VERSION_QWEN_IMAGE_LAYERED) {
+                additional_t_cond_vec.assign(static_cast<size_t>(batch_size), 0);
+                addition_t_cond = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, batch_size);
+                set_backend_tensor_data(addition_t_cond, additional_t_cond_vec.data());
+                ref_index_mode = Rope::RefIndexMode::DECREASE;
+            }
+
+            pe_vec      = Rope::gen_qwen_image_pe(time_len,
+                                                  static_cast<int>(x->ne[1]),
                                                   static_cast<int>(x->ne[0]),
                                                   config.patch_size,
-                                                  static_cast<int>(x->ne[3]),
+                                                  batch_size,
                                                   static_cast<int>(context->ne[1]),
                                                   ref_latents,
-                                                  increase_ref_index,
+                                                  ref_index_mode,
                                                   config.theta,
                                                   circular_y_enabled,
                                                   circular_x_enabled,
@@ -604,6 +669,7 @@ namespace Qwen {
             ggml_tensor* out = qwen_image.forward(&runner_ctx,
                                                   x,
                                                   timesteps,
+                                                  addition_t_cond,
                                                   context,
                                                   pe,
                                                   ref_latents,
@@ -619,12 +685,12 @@ namespace Qwen {
                                   const sd::Tensor<float>& timesteps,
                                   const sd::Tensor<float>& context,
                                   const std::vector<sd::Tensor<float>>& ref_latents = {},
-                                  bool increase_ref_index                           = false) {
-            // x: [N, in_channels, h, w]
+                                  Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::INCREASE) {
+            // x: [N, C, H, W] or [N*C, T, H, W]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+                return build_graph(x, timesteps, context, ref_latents, ref_index_mode);
             };
 
             return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
@@ -640,7 +706,7 @@ namespace Qwen {
                            *diffusion_params.timesteps,
                            tensor_or_empty(diffusion_params.context),
                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                           diffusion_params.increase_ref_index);
+                           diffusion_params.ref_index_mode);
         }
 
         void test() {
@@ -674,7 +740,7 @@ namespace Qwen {
                                        timesteps,
                                        context,
                                        {},
-                                       false);
+                                       Rope::RefIndexMode::FIXED);
                 int64_t t1   = ggml_time_ms();
 
                 GGML_ASSERT(!out_opt.empty());

src/model/diffusion/z_image.hpp

@@ -575,7 +575,7 @@ namespace ZImage {
                                  const sd::Tensor<float>& timesteps_tensor,
                                  const sd::Tensor<float>& context_tensor,
                                  const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
-                                 bool increase_ref_index                                  = false) {
+                                 Rope::RefIndexMode ref_index_mode                        = Rope::RefIndexMode::FIXED) {
             ggml_cgraph* gf        = new_graph_custom(Z_IMAGE_GRAPH_SIZE);
             ggml_tensor* x         = make_input(x_tensor);
             ggml_tensor* timesteps = make_input(timesteps_tensor);
@@ -595,7 +595,7 @@ namespace ZImage {
                                                static_cast<int>(context->ne[1]),
                                                SEQ_MULTI_OF,
                                                ref_latents,
-                                               increase_ref_index,
+                                               ref_index_mode,
                                                config.theta,
                                                circular_y_enabled,
                                                circular_x_enabled,
@@ -626,12 +626,12 @@ namespace ZImage {
                                   const sd::Tensor<float>& timesteps,
                                   const sd::Tensor<float>& context,
                                   const std::vector<sd::Tensor<float>>& ref_latents = {},
-                                  bool increase_ref_index                           = false) {
+                                  Rope::RefIndexMode ref_index_mode                 = Rope::RefIndexMode::FIXED) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+                return build_graph(x, timesteps, context, ref_latents, ref_index_mode);
             };
 
             return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
@@ -647,7 +647,7 @@ namespace ZImage {
                            *diffusion_params.timesteps,
                            tensor_or_empty(diffusion_params.context),
                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                           diffusion_params.increase_ref_index);
+                           diffusion_params.ref_index_mode);
         }
 
         void test() {
@@ -681,7 +681,7 @@ namespace ZImage {
                                        timesteps,
                                        context,
                                        {},
-                                       false);
+                                       Rope::RefIndexMode::FIXED);
                 int64_t t1   = ggml_time_ms();
 
                 GGML_ASSERT(!out_opt.empty());

src/model/te/clip.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_TE_CLIP_HPP__
+#ifndef __SD_MODEL_TE_CLIP_HPP__
 #define __SD_MODEL_TE_CLIP_HPP__
 
 #include "core/ggml_extend.hpp"

src/model/te/llm.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_TE_LLM_HPP__
+#ifndef __SD_MODEL_TE_LLM_HPP__
 #define __SD_MODEL_TE_LLM_HPP__
 
 #include <algorithm>

src/model/te/t5.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_TE_T5_HPP__
+#ifndef __SD_MODEL_TE_T5_HPP__
 #define __SD_MODEL_TE_T5_HPP__
 
 #include <cfloat>
@@ -26,13 +26,66 @@ struct T5Config {
     static T5Config detect_from_weights(const String2TensorStorage& tensor_storage_map,
                                         const std::string& prefix,
                                         bool is_umt5 = false) {
-        (void)tensor_storage_map;
-        (void)prefix;
         T5Config config;
         if (is_umt5) {
             config.vocab_size         = 256384;
             config.relative_attention = false;
         }
+        auto find_tensor = [&](const std::string& suffix) -> const TensorStorage* {
+            auto it = tensor_storage_map.find(prefix + "." + suffix);
+            if (it != tensor_storage_map.end()) {
+                return &it->second;
+            }
+            it = tensor_storage_map.find(prefix + suffix);
+            if (it != tensor_storage_map.end()) {
+                return &it->second;
+            }
+            return nullptr;
+        };
+
+        if (const TensorStorage* shared = find_tensor("shared.weight")) {
+            if (shared->n_dims == 2) {
+                config.vocab_size = shared->ne[1];
+                config.model_dim  = shared->ne[0];
+            }
+        }
+        if (const TensorStorage* q = find_tensor("encoder.block.0.layer.0.SelfAttention.q.weight")) {
+            if (q->n_dims == 2) {
+                config.model_dim  = q->ne[0];
+                int64_t inner_dim = q->ne[1];
+                // Flan-T5/T5 uses d_kv=64 for common sizes.
+                if (inner_dim % 64 == 0) {
+                    config.num_heads = inner_dim / 64;
+                }
+            }
+        }
+        if (const TensorStorage* wi = find_tensor("encoder.block.0.layer.1.DenseReluDense.wi_0.weight")) {
+            if (wi->n_dims == 2) {
+                config.model_dim = wi->ne[0];
+                config.ff_dim    = wi->ne[1];
+            }
+        }
+        int64_t detected_layers = 0;
+        for (const auto& [name, _] : tensor_storage_map) {
+            std::string base = prefix;
+            if (!base.empty() && base.back() != '.') {
+                base += ".";
+            }
+            std::string layer_prefix = base + "encoder.block.";
+            if (!starts_with(name, layer_prefix)) {
+                continue;
+            }
+            size_t pos = layer_prefix.size();
+            size_t dot = name.find('.', pos);
+            if (dot == std::string::npos) {
+                continue;
+            }
+            int64_t layer   = atoi(name.substr(pos, dot - pos).c_str());
+            detected_layers = std::max(detected_layers, layer + 1);
+        }
+        if (detected_layers > 0) {
+            config.num_layers = detected_layers;
+        }
         return config;
     }
 };

src/model/upscaler/esrgan.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_UPSCALER_ESRGAN_HPP__
+#ifndef __SD_MODEL_UPSCALER_ESRGAN_HPP__
 #define __SD_MODEL_UPSCALER_ESRGAN_HPP__
 
 #include <algorithm>

src/model/upscaler/ltx_latent_upscaler.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
+#ifndef __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
 #define __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
 
 #include <algorithm>

src/model/vae/ltx_audio_vae.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
+#ifndef __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
 #define __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
 
 #include <cmath>
@@ -214,7 +214,7 @@ namespace LTXV {
 
         auto x = ggml_reshape_3d(ctx, waveform, time, 1, channels * batch);
         if (left_pad > 0) {
-            x = ggml_pad_ext(ctx, x, static_cast<int>(left_pad), 0, 0, 0, 0, 0, 0, 0);
+            x = ggml_ext_pad_ext(ctx, runner_ctx->backend, x, static_cast<int>(left_pad), 0, 0, 0, 0, 0, 0, 0);
         }
 
         auto frames = ggml_conv_1d(ctx, forward_basis, x, hop_length, 0, 1);
@@ -451,6 +451,7 @@ namespace LTXV {
             int pad_h = kernel_size.first - 1;
             int pad_w = kernel_size.second - 1;
             x         = ggml_ext_pad_ext(ctx->ggml_ctx,
+                                         ctx->backend,
                                          x,
                                          pad_w / 2,
                                          pad_w - pad_w / 2,

src/model/vae/tae.hpp

@@ -408,7 +408,7 @@ public:
             h = conv->forward(ctx, h);
             for (int j = 0; j < num_blocks; j++) {
                 auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
-                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                auto mem   = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, h, 0, 0, 0, 0, 0, 0, 1, 0);
                 mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
                 h          = block->forward(ctx, h, mem);
             }
@@ -479,7 +479,7 @@ public:
         int index = 3;
         for (int i = 0; i < num_layers; i++) {
             for (int j = 0; j < num_blocks; j++) {
-                auto mem = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                auto mem = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, h, 0, 0, 0, 0, 0, 0, 1, 0);
                 mem      = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
                 if (is_wide) {
                     auto block = std::dynamic_pointer_cast<WideMemBlock>(blocks[std::to_string(index++)]);

src/model/vae/vae.hpp

@@ -1,4 +1,4 @@
-#ifndef __SD_MODEL_VAE_VAE_HPP__
+#ifndef __SD_MODEL_VAE_VAE_HPP__
 #define __SD_MODEL_VAE_VAE_HPP__
 
 #include "core/tensor_ggml.hpp"
@@ -78,7 +78,7 @@ 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 || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version)) {
             scale_factor = 1;
         }
         return scale_factor;
@@ -133,7 +133,11 @@ public:
             int64_t H              = input.shape()[1] / scale_factor;
             float tile_overlap;
             int tile_size_x, tile_size_y;
-            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, 1.30539f);
+            // Image VAE encode is more sensitive to tile boundary context than decode.
+            // Keep the smaller legacy factor for video VAEs, but default image encode
+            // tiles to 64 latent pixels so a 512px SD image is encoded as one tile.
+            const float encode_tile_factor = (sd_version_is_wan(version) || sd_version_is_ltxav(version)) ? 1.30539f : 2.0f;
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, encode_tile_factor);
             LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
             output = tiled_compute(input,
                                    n_threads,

src/model/vae/wan_vae.hpp

@@ -72,8 +72,8 @@ namespace WAN {
                 lp2 -= (int)cache_x->ne[2];
             }
 
-            x = ggml_ext_pad_ext(ctx->ggml_ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
-            return ggml_ext_conv_3d(ctx->ggml_ctx, x, w, b, in_channels,
+            x = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, x, lp0, rp0, lp1, rp1, lp2, rp2, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
+            return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
                                     std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                     0, 0, 0,
                                     std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation));
@@ -195,7 +195,7 @@ namespace WAN {
                                                   2);
                         }
                         if (chunk_idx == 1 && cache_x->ne[2] < 2) {  // Rep
-                            cache_x = ggml_pad_ext(ctx->ggml_ctx, cache_x, 0, 0, 0, 0, (int)cache_x->ne[2], 0, 0, 0);
+                            cache_x = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, cache_x, 0, 0, 0, 0, (int)cache_x->ne[2], 0, 0, 0);
                             // aka cache_x = torch.cat([torch.zeros_like(cache_x).to(cache_x.device),cache_x],dim=2)
                         }
                         if (chunk_idx == 1) {
@@ -283,7 +283,7 @@ namespace WAN {
 
             int pad_t = (factor_t - T % factor_t) % factor_t;
 
-            x = ggml_pad_ext(ctx->ggml_ctx, x, 0, 0, 0, 0, pad_t, 0, 0, 0);
+            x = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, x, 0, 0, 0, 0, pad_t, 0, 0, 0);
             T = x->ne[2];
 
             x = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, factor_t, T / factor_t, C);                                                  // [C, T/factor_t, factor_t, H*W]
@@ -631,6 +631,7 @@ namespace WAN {
     class Encoder3d : public GGMLBlock {
     protected:
         bool wan2_2;
+        int64_t in_channels;
         int64_t dim;
         int64_t z_dim;
         std::vector<int> dim_mult;
@@ -641,12 +642,14 @@ namespace WAN {
     public:
         Encoder3d(int64_t dim                           = 128,
                   int64_t z_dim                         = 4,
+                  int64_t in_channels                   = 3,
                   std::vector<int> dim_mult             = {1, 2, 4, 4},
                   int num_res_blocks                    = 2,
                   std::vector<bool> temperal_downsample = {false, true, true},
                   bool wan2_2                           = false,
                   bool is_2D                            = false)
-            : dim(dim),
+            : in_channels(in_channels),
+              dim(dim),
               z_dim(z_dim),
               dim_mult(dim_mult),
               num_res_blocks(num_res_blocks),
@@ -659,11 +662,10 @@ namespace WAN {
                 dims.push_back(dim * u);
             }
 
-            int64_t input_dim = wan2_2 ? 12 : 3;
             if (is_2D) {
-                blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2dBut3d(input_dim, dims[0], {3, 3}, {1, 1}, {1, 1}));
+                blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2dBut3d(in_channels, dims[0], {3, 3}, {1, 1}, {1, 1}));
             } else {
-                blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(input_dim, dims[0], {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
+                blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(in_channels, dims[0], {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
             }
 
             int index = 0;
@@ -812,6 +814,7 @@ namespace WAN {
     class Decoder3d : public GGMLBlock {
     protected:
         bool wan2_2;
+        int64_t out_channels;
         int64_t dim;
         int64_t z_dim;
         std::vector<int> dim_mult;
@@ -822,12 +825,14 @@ namespace WAN {
     public:
         Decoder3d(int64_t dim                         = 128,
                   int64_t z_dim                       = 4,
+                  int64_t out_channels                = 3,
                   std::vector<int> dim_mult           = {1, 2, 4, 4},
                   int num_res_blocks                  = 2,
                   std::vector<bool> temperal_upsample = {true, true, false},
                   bool wan2_2                         = false,
                   bool is_2D                          = false)
-            : dim(dim),
+            : out_channels(out_channels),
+              dim(dim),
               z_dim(z_dim),
               dim_mult(dim_mult),
               num_res_blocks(num_res_blocks),
@@ -889,7 +894,7 @@ namespace WAN {
 
             // output blocks
             blocks["head.0"]  = std::shared_ptr<GGMLBlock>(new RMS_norm(out_dim));
-            int64_t final_dim = wan2_2 ? 12 : 3;
+            int64_t final_dim = out_channels;
             // head.1 is nn.SiLU()
             if (is_2D) {
                 blocks["head.2"] = std::shared_ptr<GGMLBlock>(new Conv2dBut3d(out_dim, final_dim, {3, 3}, {1, 1}, {1, 1}));
@@ -1002,6 +1007,8 @@ namespace WAN {
     public:
         bool wan2_2                           = false;
         bool decode_only                      = true;
+        int64_t input_channels                = 3;
+        int patch_size                        = 1;
         int64_t dim                           = 96;
         int64_t dec_dim                       = 96;
         int64_t z_dim                         = 16;
@@ -1026,16 +1033,22 @@ namespace WAN {
         }
 
     public:
-        WanVAE(bool decode_only = true, bool wan2_2 = false, bool is_2D = false)
-            : decode_only(decode_only), wan2_2(wan2_2), is_2D(is_2D) {
+        WanVAE(bool decode_only = true, SDVersion version = VERSION_WAN2, bool is_2D = false)
+            : decode_only(decode_only),
+              wan2_2(version == VERSION_WAN2_2_TI2V),
+              is_2D(is_2D) {
             // attn_scales is always []
             if (wan2_2) {
-                dim     = 160;
-                dec_dim = 256;
-                z_dim   = 48;
+                dim            = 160;
+                dec_dim        = 256;
+                z_dim          = 48;
+                input_channels = 12;
+                patch_size     = 2;
 
                 _conv_num     = 34;
                 _enc_conv_num = 26;
+            } else if (version == VERSION_QWEN_IMAGE_LAYERED) {
+                input_channels = 4;
             }
 
             if (is_2D) {
@@ -1044,14 +1057,14 @@ namespace WAN {
             }
 
             if (!decode_only) {
-                blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks, temperal_downsample, wan2_2, is_2D));
+                blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder3d(dim, z_dim * 2, input_channels, dim_mult, num_res_blocks, temperal_downsample, wan2_2, is_2D));
                 if (is_2D) {
                     blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2dBut3d(z_dim * 2, z_dim * 2, {1, 1}));
                 } else {
                     blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim * 2, z_dim * 2, {1, 1, 1}));
                 }
             }
-            blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder3d(dec_dim, z_dim, dim_mult, num_res_blocks, temperal_upsample, wan2_2, is_2D));
+            blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder3d(dec_dim, z_dim, input_channels, dim_mult, num_res_blocks, temperal_upsample, wan2_2, is_2D));
             if (is_2D) {
                 blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2dBut3d(z_dim, z_dim, {1, 1}));
             } else {
@@ -1125,9 +1138,7 @@ namespace WAN {
 
             clear_cache();
 
-            if (wan2_2) {
-                x = patchify(ctx->ggml_ctx, x, 2, b);
-            }
+            x = patchify(ctx->ggml_ctx, x, patch_size, b);
             // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encode.prelude", "x");
 
             auto encoder = std::dynamic_pointer_cast<Encoder3d>(blocks["encoder"]);
@@ -1202,9 +1213,7 @@ namespace WAN {
                     }
                 }
             }
-            if (wan2_2) {
-                out = unpatchify(ctx->ggml_ctx, out, 2, b);
-            }
+            out = unpatchify(ctx->ggml_ctx, out, patch_size, b);
             // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode.final", "out");
             clear_cache();
             return out;
@@ -1225,9 +1234,7 @@ namespace WAN {
             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);
-            }
+            out       = unpatchify(ctx->ggml_ctx, out, patch_size, b);
             // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode_partial.final", "out");
             return out;
         }
@@ -1257,7 +1264,7 @@ namespace WAN {
             if (is_2D) {
                 LOG_DEBUG("USING 2D VAE");
             }
-            ae = WanVAE(decode_only, version == VERSION_WAN2_2_TI2V, is_2D);
+            ae = WanVAE(decode_only, version, is_2D);
             ae.init(params_ctx, tensor_storage_map, prefix);
         }
 

src/model_io/gguf_io.cpp

@@ -1,7 +1,10 @@
 #include "gguf_io.h"
 
+#include <algorithm>
 #include <cstdint>
+#include <cstdio>
 #include <fstream>
+#include <ostream>
 #include <string>
 #include <vector>
 
@@ -121,3 +124,115 @@ bool write_gguf_file(const std::string& file_path,
     gguf_free(gguf_ctx);
     return success;
 }
+
+GGUFStreamingWriter::~GGUFStreamingWriter() {
+    close();
+}
+
+bool GGUFStreamingWriter::write_metadata(const std::string& file_path,
+                                         const std::vector<TensorWritePlan>& tensors,
+                                         std::string* error) {
+    close();
+    tensors_   = tensors;
+    file_size_ = 0;
+
+    size_t meta_mem = 1 * 1024 * 1024 + tensors.size() * ggml_tensor_overhead();
+    meta_ctx_       = ggml_init({meta_mem, nullptr, true});
+    if (meta_ctx_ == nullptr) {
+        set_error(error, "ggml_init failed for GGUF metadata");
+        return false;
+    }
+
+    gguf_ctx_ = gguf_init_empty();
+    if (gguf_ctx_ == nullptr) {
+        set_error(error, "gguf_init_empty failed");
+        close();
+        return false;
+    }
+
+    for (const TensorWritePlan& plan : tensors) {
+        ggml_tensor* tensor = ggml_new_tensor(meta_ctx_, plan.type, plan.n_dims, plan.ne);
+        if (tensor == nullptr) {
+            set_error(error, "ggml_new_tensor failed for tensor '" + plan.name + "'");
+            close();
+            return false;
+        }
+        ggml_set_name(tensor, plan.name.c_str());
+        gguf_add_tensor(gguf_ctx_, tensor);
+    }
+
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    FILE* file = fopen(file_path.c_str(), "wb+");
+    if (file == nullptr) {
+        set_error(error, "failed to open output file '" + file_path + "'");
+        close();
+        return false;
+    }
+
+    // ggml exposes GGUF metadata writing through FILE* only. Keep FILE usage
+    // isolated here; tensor data is written through std::fstream by the shared
+    // streaming pipeline.
+    if (!gguf_write_to_file_ptr(gguf_ctx_, file, true)) {
+        fclose(file);
+        set_error(error, "failed to write GGUF metadata to '" + file_path + "'");
+        close();
+        return false;
+    }
+    fclose(file);
+
+    const uint64_t data_start = gguf_get_meta_size(gguf_ctx_);
+    tensor_offsets_.resize(tensors.size());
+    file_size_ = data_start;
+    for (size_t i = 0; i < tensors.size(); i++) {
+        tensor_offsets_[i] = data_start + gguf_get_tensor_offset(gguf_ctx_, static_cast<int64_t>(i));
+        file_size_         = std::max(file_size_, tensor_offsets_[i] + tensors[i].nbytes());
+    }
+    return true;
+}
+
+bool GGUFStreamingWriter::write_tensor(std::ostream& output,
+                                       size_t tensor_index,
+                                       const uint8_t* data,
+                                       size_t size,
+                                       std::string* error) const {
+    if (tensor_index >= tensors_.size() || tensor_index >= tensor_offsets_.size()) {
+        set_error(error, "invalid GGUF tensor index");
+        return false;
+    }
+    const TensorWritePlan& plan = tensors_[tensor_index];
+    if (size != plan.nbytes()) {
+        set_error(error, "size mismatch while writing tensor '" + plan.name + "'");
+        return false;
+    }
+    output.seekp(static_cast<std::streamoff>(tensor_offsets_[tensor_index]), std::ios::beg);
+    if (!output) {
+        set_error(error, "failed to seek output for tensor '" + plan.name + "'");
+        return false;
+    }
+    if (size > 0) {
+        output.write(reinterpret_cast<const char*>(data), static_cast<std::streamsize>(size));
+    }
+    if (!output) {
+        set_error(error, "failed to write tensor '" + plan.name + "'");
+        return false;
+    }
+    return true;
+}
+
+uint64_t GGUFStreamingWriter::file_size() const {
+    return file_size_;
+}
+
+void GGUFStreamingWriter::close() {
+    tensor_offsets_.clear();
+    tensors_.clear();
+    file_size_ = 0;
+    if (gguf_ctx_ != nullptr) {
+        gguf_free(gguf_ctx_);
+        gguf_ctx_ = nullptr;
+    }
+    if (meta_ctx_ != nullptr) {
+        ggml_free(meta_ctx_);
+        meta_ctx_ = nullptr;
+    }
+}

src/model_io/gguf_io.h

@@ -4,8 +4,12 @@
 #include <string>
 #include <vector>
 
+#include "streaming_writer.h"
 #include "tensor_storage.h"
 
+struct ggml_context;
+struct gguf_context;
+
 bool is_gguf_file(const std::string& file_path);
 bool read_gguf_file(const std::string& file_path,
                     std::vector<TensorStorage>& tensor_storages,
@@ -14,4 +18,28 @@ bool write_gguf_file(const std::string& file_path,
                      const std::vector<TensorWriteInfo>& tensors,
                      std::string* error = nullptr);
 
+class GGUFStreamingWriter : public StreamingModelWriter {
+public:
+    GGUFStreamingWriter() = default;
+    ~GGUFStreamingWriter();
+
+    bool write_metadata(const std::string& file_path,
+                        const std::vector<TensorWritePlan>& tensors,
+                        std::string* error = nullptr) override;
+    bool write_tensor(std::ostream& output,
+                      size_t tensor_index,
+                      const uint8_t* data,
+                      size_t size,
+                      std::string* error = nullptr) const override;
+    uint64_t file_size() const override;
+    void close();
+
+private:
+    std::vector<TensorWritePlan> tensors_;
+    std::vector<uint64_t> tensor_offsets_;
+    uint64_t file_size_     = 0;
+    ggml_context* meta_ctx_ = nullptr;
+    gguf_context* gguf_ctx_ = nullptr;
+};
+
 #endif  // __SD_MODEL_IO_GGUF_IO_H__

src/model_io/safetensors_io.cpp

@@ -1,8 +1,10 @@
 #include "safetensors_io.h"
 
+#include <algorithm>
 #include <cstdint>
 #include <exception>
 #include <fstream>
+#include <ostream>
 #include <string>
 #include <vector>
 
@@ -314,3 +316,102 @@ bool write_safetensors_file(const std::string& file_path,
 
     return true;
 }
+
+bool SafetensorsStreamingWriter::write_metadata(const std::string& file_path,
+                                                const std::vector<TensorWritePlan>& tensors,
+                                                std::string* error) {
+    file_path_ = file_path;
+    tensors_   = tensors;
+    tensor_offsets_.clear();
+    data_start_ = 0;
+    file_size_  = 0;
+
+    nlohmann::ordered_json header = nlohmann::ordered_json::object();
+    uint64_t data_offset          = 0;
+    tensor_offsets_.resize(tensors.size());
+    for (size_t i = 0; i < tensors.size(); i++) {
+        const TensorWritePlan& plan = tensors[i];
+        std::string dtype;
+        if (!ggml_type_to_safetensors_dtype(plan.type, &dtype)) {
+            set_error(error,
+                      "unsupported safetensors dtype '" + std::string(ggml_type_name(plan.type)) +
+                          "' for tensor '" + plan.name + "'");
+            return false;
+        }
+
+        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 j = 0; j < plan.n_dims; ++j) {
+            shape.push_back(plan.ne[plan.n_dims - 1 - j]);
+        }
+        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 + plan.nbytes());
+        json_tensor_info["data_offsets"] = data_offsets;
+
+        header[plan.name]  = json_tensor_info;
+        tensor_offsets_[i] = data_offset;
+        data_offset += plan.nbytes();
+    }
+
+    const std::string header_str = header.dump();
+    data_start_                  = ST_HEADER_SIZE_LEN + header_str.size();
+
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    std::ofstream file(file_path, std::ios::binary | std::ios::trunc);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "' for writing");
+        return false;
+    }
+
+    uint8_t header_size[ST_HEADER_SIZE_LEN];
+    for (int i = 0; i < static_cast<int>(ST_HEADER_SIZE_LEN); ++i) {
+        header_size[i] = static_cast<uint8_t>((header_str.size() >> (8 * i)) & 0xFF);
+    }
+    file.write(reinterpret_cast<const char*>(header_size), sizeof(header_size));
+    file.write(header_str.data(), static_cast<std::streamsize>(header_str.size()));
+    if (!file) {
+        set_error(error, "failed to write safetensors header to '" + file_path + "'");
+        return false;
+    }
+
+    file_size_ = data_start_ + data_offset;
+    return true;
+}
+
+bool SafetensorsStreamingWriter::write_tensor(std::ostream& output,
+                                              size_t tensor_index,
+                                              const uint8_t* data,
+                                              size_t size,
+                                              std::string* error) const {
+    if (tensor_index >= tensors_.size() || tensor_index >= tensor_offsets_.size()) {
+        set_error(error, "invalid safetensors tensor index");
+        return false;
+    }
+    const TensorWritePlan& plan = tensors_[tensor_index];
+    if (size != plan.nbytes()) {
+        set_error(error, "size mismatch while writing tensor '" + plan.name + "'");
+        return false;
+    }
+    output.seekp(static_cast<std::streamoff>(data_start_ + tensor_offsets_[tensor_index]), std::ios::beg);
+    if (!output) {
+        set_error(error, "failed to seek output for tensor '" + plan.name + "'");
+        return false;
+    }
+    if (size > 0) {
+        output.write(reinterpret_cast<const char*>(data), static_cast<std::streamsize>(size));
+    }
+    if (!output) {
+        set_error(error, "failed to write tensor '" + plan.name + "' to '" + file_path_ + "'");
+        return false;
+    }
+    return true;
+}
+
+uint64_t SafetensorsStreamingWriter::file_size() const {
+    return file_size_;
+}

src/model_io/safetensors_io.h

@@ -4,6 +4,7 @@
 #include <string>
 #include <vector>
 
+#include "streaming_writer.h"
 #include "tensor_storage.h"
 
 bool is_safetensors_file(const std::string& file_path);
@@ -14,4 +15,26 @@ bool write_safetensors_file(const std::string& file_path,
                             const std::vector<TensorWriteInfo>& tensors,
                             std::string* error = nullptr);
 
+class SafetensorsStreamingWriter : public StreamingModelWriter {
+public:
+    SafetensorsStreamingWriter() = default;
+
+    bool write_metadata(const std::string& file_path,
+                        const std::vector<TensorWritePlan>& tensors,
+                        std::string* error = nullptr) override;
+    bool write_tensor(std::ostream& output,
+                      size_t tensor_index,
+                      const uint8_t* data,
+                      size_t size,
+                      std::string* error = nullptr) const override;
+    uint64_t file_size() const override;
+
+private:
+    std::string file_path_;
+    std::vector<TensorWritePlan> tensors_;
+    std::vector<uint64_t> tensor_offsets_;
+    uint64_t data_start_ = 0;
+    uint64_t file_size_  = 0;
+};
+
 #endif  // __SD_MODEL_IO_SAFETENSORS_IO_H__

src/model_io/streaming_writer.h

@@ -0,0 +1,26 @@
+#ifndef __SD_MODEL_IO_STREAMING_WRITER_H__
+#define __SD_MODEL_IO_STREAMING_WRITER_H__
+
+#include <cstdint>
+#include <iosfwd>
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+class StreamingModelWriter {
+public:
+    virtual ~StreamingModelWriter() = default;
+
+    virtual bool write_metadata(const std::string& file_path,
+                                const std::vector<TensorWritePlan>& tensors,
+                                std::string* error = nullptr)     = 0;
+    virtual bool write_tensor(std::ostream& output,
+                              size_t tensor_index,
+                              const uint8_t* data,
+                              size_t size,
+                              std::string* error = nullptr) const = 0;
+    virtual uint64_t file_size() const                            = 0;
+};
+
+#endif  // __SD_MODEL_IO_STREAMING_WRITER_H__

src/model_io/tensor_storage.h

@@ -127,6 +127,25 @@ struct TensorWriteInfo {
     ggml_tensor* tensor     = nullptr;
 };
 
+struct TensorWritePlan {
+    std::string name;
+    ggml_type type          = GGML_TYPE_F32;
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
+
+    int64_t nelements() const {
+        int64_t n = 1;
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            n *= ne[i];
+        }
+        return n;
+    }
+
+    uint64_t nbytes() const {
+        return nelements() * ggml_type_size(type) / ggml_blck_size(type);
+    }
+};
+
 typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
 
 #endif  // __SD_TENSOR_STORAGE_H__

src/model_loader.cpp

@@ -20,6 +20,7 @@
 #include "model_io/torch_legacy_io.h"
 #include "model_io/torch_zip_io.h"
 #include "model_loader.h"
+#include "runtime/imatrix.h"
 #include "stable-diffusion.h"
 
 #include "core/ggml_extend_backend.h"
@@ -156,7 +157,8 @@ void convert_tensor(void* src,
                     void* dst,
                     ggml_type dst_type,
                     int nrows,
-                    int n_per_row) {
+                    int n_per_row,
+                    std::vector<float> imatrix = {}) {
     int n = nrows * n_per_row;
     if (src_type == dst_type) {
         size_t nbytes = n * ggml_type_size(src_type) / ggml_blck_size(src_type);
@@ -165,7 +167,7 @@ void convert_tensor(void* src,
         if (dst_type == GGML_TYPE_F16) {
             ggml_fp32_to_fp16_row((float*)src, (ggml_fp16_t*)dst, n);
         } else {
-            std::vector<float> imatrix(n_per_row, 1.0f);  // dummy importance matrix
+            imatrix.resize(n_per_row, 1.0f);
             const float* im = imatrix.data();
             ggml_quantize_chunk(dst_type, (float*)src, dst, 0, nrows, n_per_row, im);
         }
@@ -195,7 +197,7 @@ void convert_tensor(void* src,
         if (dst_type == GGML_TYPE_F16) {
             ggml_fp32_to_fp16_row((float*)src_data_f32, (ggml_fp16_t*)dst, n);
         } else {
-            std::vector<float> imatrix(n_per_row, 1.0f);  // dummy importance matrix
+            imatrix.resize(n_per_row, 1.0f);
             const float* im = imatrix.data();
             ggml_quantize_chunk(dst_type, (float*)src_data_f32, dst, 0, nrows, n_per_row, im);
         }
@@ -470,7 +472,13 @@ SDVersion ModelLoader::get_sd_version() {
             tensor_storage_map.find("model.diffusion_model.transformer_blocks.0.img_mlp.w1.weight") != tensor_storage_map.end()) {
             return VERSION_LENS;
         }
+        if (tensor_storage.name.find("net.img_embedder.proj1.weight") != std::string::npos) {
+            return VERSION_MINIT2I;
+        }
         if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
+            if (tensor_storage_map.find("model.diffusion_model.time_text_embed.addition_t_embedding.weight") != tensor_storage_map.end()) {
+                return VERSION_QWEN_IMAGE_LAYERED;
+            }
             return VERSION_QWEN_IMAGE;
         }
         if (tensor_storage.name.find("llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
@@ -935,7 +943,8 @@ std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggm
 
 bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
                                bool enable_mmap,
-                               const std::set<std::string>* target_tensor_names) {
+                               const std::set<std::string>* target_tensor_names,
+                               bool log_progress) {
     process_model_files(enable_mmap, false);
 
     std::atomic<int64_t> read_time_ms(0);
@@ -970,6 +979,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
     size_t total_tensors_processed = 0;
     const int64_t t_start          = start_time;
     int last_n_threads             = 1;
+    SDVersion imatrix_version      = (version_ == VERSION_COUNT) ? get_sd_version() : version_;
 
     for (size_t file_index = 0; file_index < file_data.size(); ++file_index) {
         auto& fdata                  = file_data[file_index];
@@ -1154,12 +1164,15 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
                             failed = true;
                             return;
                         }
+                        std::string processed_name = convert_tensor_name(tensor_storage.name, imatrix_version);
+                        std::vector<float> imatrix = get_imatrix_collector().get_values(processed_name);
                         convert_tensor((void*)target_buf,
                                        tensor_storage.type,
                                        convert_buf,
                                        dst_tensor->type,
                                        (int)tensor_storage.nelements() / (int)tensor_storage.ne[0],
-                                       (int)tensor_storage.ne[0]);
+                                       (int)tensor_storage.ne[0],
+                                       std::move(imatrix));
                     } else {
                         convert_buf = read_buf;
                     }
@@ -1200,7 +1213,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
             }
             size_t curr_num       = total_tensors_processed + current_idx;
             float elapsed_seconds = (ggml_time_ms() - t_start) / 1000.0f;
-            if (total_tensors_to_process > 0) {
+            if (log_progress && total_tensors_to_process > 0) {
                 pretty_bytes_progress(static_cast<int>(curr_num),
                                       static_cast<int>(total_tensors_to_process),
                                       bytes_processed.load(),
@@ -1218,27 +1231,81 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
             break;
         }
         total_tensors_processed += tensors_to_process.size();
-        if (total_tensors_to_process > 0) {
+        if (log_progress && total_tensors_to_process > 0) {
             pretty_bytes_progress(static_cast<int>(total_tensors_processed),
                                   static_cast<int>(total_tensors_to_process),
                                   bytes_processed.load(),
                                   (ggml_time_ms() - t_start) / 1000.0f);
         }
-        if (total_tensors_processed < total_tensors_to_process && total_tensors_to_process > 0) {
+        if (log_progress && total_tensors_processed < total_tensors_to_process && total_tensors_to_process > 0) {
             printf("\n");
         }
     }
 
     int64_t end_time = ggml_time_ms();
-    LOG_INFO("loading tensors completed, taking %.2fs (read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
-             (end_time - start_time) / 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,
-             (copy_to_backend_time_ms.load() / (float)last_n_threads) / 1000.f);
+    if (log_progress) {
+        LOG_INFO("loading tensors completed, taking %.2fs (read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
+                 (end_time - start_time) / 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,
+                 (copy_to_backend_time_ms.load() / (float)last_n_threads) / 1000.f);
+    }
     return success;
 }
 
+bool ModelLoader::load_tensor(const TensorStorage& tensor_storage, ggml_tensor* dst_tensor) {
+    if (dst_tensor == nullptr || dst_tensor->data == nullptr) {
+        LOG_ERROR("load tensor failed: null destination for '%s'", tensor_storage.name.c_str());
+        return false;
+    }
+
+    bool loaded = false;
+    std::set<std::string> target_tensor_names{tensor_storage.name};
+    auto on_new_tensor_cb = [&](const TensorStorage& current_tensor_storage, ggml_tensor** out_tensor) -> bool {
+        *out_tensor = nullptr;
+        if (current_tensor_storage.name != tensor_storage.name) {
+            return true;
+        }
+
+        if (current_tensor_storage.file_index != tensor_storage.file_index ||
+            current_tensor_storage.offset != tensor_storage.offset ||
+            current_tensor_storage.index_in_zip != tensor_storage.index_in_zip) {
+            LOG_ERROR("load tensor failed: storage mismatch for '%s'", tensor_storage.name.c_str());
+            return false;
+        }
+
+        if (current_tensor_storage.n_dims != tensor_storage.n_dims ||
+            current_tensor_storage.nelements() != tensor_storage.nelements()) {
+            LOG_ERROR("load tensor failed: metadata changed for '%s'", tensor_storage.name.c_str());
+            return false;
+        }
+
+        for (int i = 0; i < current_tensor_storage.n_dims; i++) {
+            if (current_tensor_storage.ne[i] != dst_tensor->ne[i]) {
+                LOG_ERROR("load tensor failed: shape mismatch for '%s'", tensor_storage.name.c_str());
+                return false;
+            }
+        }
+
+        *out_tensor = dst_tensor;
+        loaded      = true;
+        return true;
+    };
+
+    if (!load_tensors(on_new_tensor_cb, false, &target_tensor_names, false)) {
+        LOG_ERROR("load tensor failed: '%s'", tensor_storage.name.c_str());
+        return false;
+    }
+
+    if (!loaded) {
+        LOG_ERROR("load tensor failed: tensor '%s' not found", tensor_storage.name.c_str());
+        return false;
+    }
+
+    return true;
+}
+
 bool ModelLoader::load_float_tensor(const std::string& name,
                                     std::vector<float>& data,
                                     int n_threads,

src/model_loader.h

@@ -27,6 +27,8 @@ struct MmapTensorStore {
     std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
 };
 
+bool is_unused_tensor(const std::string& name);
+
 class ModelLoader {
 protected:
     SDVersion version_ = VERSION_COUNT;
@@ -68,7 +70,8 @@ public:
                                               bool writable                        = true);
     bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
                       bool use_mmap                                    = false,
-                      const std::set<std::string>* target_tensor_names = nullptr);
+                      const std::set<std::string>* target_tensor_names = nullptr,
+                      bool log_progress                                = true);
     bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                       std::set<std::string> ignore_tensors = {},
                       bool use_mmap                        = false);
@@ -76,6 +79,7 @@ public:
                            std::vector<float>& data,
                            int n_threads = 0,
                            bool use_mmap = false);
+    bool load_tensor(const TensorStorage& tensor_storage, ggml_tensor* dst_tensor);
 
     std::vector<std::string> get_tensor_names() const {
         std::vector<std::string> names;

src/model_manager.cpp

@@ -100,12 +100,41 @@ size_t estimate_tensors_size(const std::map<std::string, ggml_tensor*>& tensors)
     return size;
 }
 
+void ModelManager::set_split_buffer_type(ggml_backend_t compute_backend, ggml_backend_buffer_type_t split_buft) {
+    if (compute_backend == nullptr) {
+        return;
+    }
+    if (split_buft == nullptr) {
+        split_buffer_types_.erase(compute_backend);
+        return;
+    }
+    split_buffer_types_[compute_backend] = split_buft;
+}
+
+bool ModelManager::tensor_shape_supports_split_buffer(const ggml_tensor* tensor) {
+    return tensor != nullptr &&
+           tensor->view_src == nullptr &&
+           ggml_is_contiguous(tensor) &&
+           ggml_n_dims(tensor) == 2 &&
+           tensor->ne[0] >= 256 &&
+           tensor->ne[1] >= 256;
+}
+
+ggml_backend_buffer_type_t ModelManager::split_buffer_type_for(const TensorState& state) const {
+    if (!state.allow_split_buffer || !tensor_shape_supports_split_buffer(state.tensor)) {
+        return nullptr;
+    }
+    auto it = split_buffer_types_.find(state.compute_backend);
+    return it != split_buffer_types_.end() ? it->second : nullptr;
+}
+
 bool ModelManager::register_param_tensors(const std::string& desc,
                                           std::map<std::string, ggml_tensor*> tensors,
                                           ResidencyMode residency_mode,
                                           ggml_backend_t compute_backend,
                                           ggml_backend_t params_backend,
-                                          size_t* registered_tensor_size) {
+                                          size_t* registered_tensor_size,
+                                          bool allow_split_buffer) {
     if (desc.empty()) {
         LOG_ERROR("model manager tensor desc is empty");
         return false;
@@ -129,13 +158,14 @@ bool ModelManager::register_param_tensors(const std::string& desc,
         }
         ggml_set_name(tensor, name.c_str());
 
-        auto state             = std::make_unique<TensorState>();
-        state->name            = name;
-        state->tensor          = tensor;
-        state->desc            = desc;
-        state->residency_mode  = residency_mode;
-        state->compute_backend = compute_backend;
-        state->params_backend  = params_backend;
+        auto state                = std::make_unique<TensorState>();
+        state->name               = name;
+        state->tensor             = tensor;
+        state->desc               = desc;
+        state->residency_mode     = residency_mode;
+        state->compute_backend    = compute_backend;
+        state->params_backend     = params_backend;
+        state->allow_split_buffer = allow_split_buffer;
         new_states.push_back(std::move(state));
     }
 
@@ -237,7 +267,7 @@ bool ModelManager::load_tensors_to_params_backend(const std::vector<TensorState*
 }
 
 bool ModelManager::stage_tensors_to_compute_backend(const std::vector<TensorState*>& states) {
-    std::map<ggml_backend_t, std::vector<TensorState*>> states_by_compute_backend;
+    std::map<std::pair<ggml_backend_t, ggml_backend_buffer_type_t>, std::vector<TensorState*>> states_by_staging_target;
     for (TensorState* state : states) {
         if (state == nullptr || should_ignore(*state) || is_optional_missing_tensor(state->name)) {
             continue;
@@ -257,11 +287,16 @@ bool ModelManager::stage_tensors_to_compute_backend(const std::vector<TensorStat
             LOG_ERROR("model manager tensor '%s' is not loaded to params backend", state->name.c_str());
             return false;
         }
-        states_by_compute_backend[state->compute_backend].push_back(state);
+        ggml_backend_buffer_type_t staging_buft = split_buffer_type_for(*state);
+        if (staging_buft == nullptr) {
+            staging_buft = ggml_backend_get_default_buffer_type(state->compute_backend);
+        }
+        states_by_staging_target[{state->compute_backend, staging_buft}].push_back(state);
     }
 
-    for (const auto& pair : states_by_compute_backend) {
-        ggml_backend_t compute_backend          = pair.first;
+    for (const auto& pair : states_by_staging_target) {
+        ggml_backend_t compute_backend          = pair.first.first;
+        ggml_backend_buffer_type_t staging_buft = pair.first.second;
         const std::vector<TensorState*>& states = pair.second;
         if (states.empty()) {
             continue;
@@ -285,7 +320,7 @@ bool ModelManager::stage_tensors_to_compute_backend(const std::vector<TensorStat
             staged_tensors.push_back({state, staging_tensor});
         }
 
-        ggml_backend_buffer_t compute_buffer = ggml_backend_alloc_ctx_tensors(staging_ctx, compute_backend);
+        ggml_backend_buffer_t compute_buffer = ggml_backend_alloc_ctx_tensors_from_buft(staging_ctx, staging_buft);
         if (compute_buffer == nullptr) {
             LOG_ERROR("model manager alloc compute params backend buffer failed, num_tensors = %zu",
                       staged_tensors.size());
@@ -350,6 +385,17 @@ bool ModelManager::apply_loras_to_params(const std::vector<TensorState*>& states
             LOG_ERROR("model manager compute backend is null for lora target tensor '%s'", state->name.c_str());
             return false;
         }
+        if (state->tensor->buffer != nullptr &&
+            ggml_backend_buffer_get_type(state->tensor->buffer) == split_buffer_type_for(*state)) {
+            if (!warned_split_lora_skip_) {
+                LOG_WARN(
+                    "model manager skipping direct lora application to row-split tensors "
+                    "(use --lora-apply-mode at_runtime with row split)");
+                warned_split_lora_skip_ = true;
+            }
+            state->applied_lora_epoch = current_lora_epoch_;
+            continue;
+        }
         if (state->tensor->data == nullptr) {
             LOG_ERROR("model manager lora target tensor '%s' is not prepared", state->name.c_str());
             return false;
@@ -694,6 +740,8 @@ ggml_backend_buffer_type_t ModelManager::params_buffer_type_for(const TensorStat
         if (compute_dev != nullptr) {
             params_buft = ggml_backend_dev_host_buffer_type(compute_dev);
         }
+    } else if (state.params_backend == state.compute_backend) {
+        params_buft = split_buffer_type_for(state);
     }
     if (params_buft == nullptr) {
         params_buft = ggml_backend_get_default_buffer_type(state.params_backend);

src/model_manager.h

@@ -36,6 +36,7 @@ private:
         ResidencyMode residency_mode   = ResidencyMode::ParamBackend;
         ggml_backend_t compute_backend = nullptr;
         ggml_backend_t params_backend  = nullptr;
+        bool allow_split_buffer        = false;
         bool metadata_validated        = false;
 
         int active_prepare_count = 0;
@@ -63,6 +64,8 @@ private:
     std::map<std::string, TensorState*> tensor_states_by_name_;
     std::vector<std::unique_ptr<ParamsStorageBlock>> params_storage_blocks_;
     std::vector<std::unique_ptr<ComputeStagingBlock>> compute_staging_blocks_;
+    std::map<ggml_backend_t, ggml_backend_buffer_type_t> split_buffer_types_;
+    bool warned_split_lora_skip_ = false;
     std::set<std::string> common_ignore_tensors_;
     std::vector<LoraSpec> loras_;
     SDVersion lora_version_      = VERSION_COUNT;
@@ -91,6 +94,7 @@ private:
     bool stage_tensors_to_compute_backend(const std::vector<TensorState*>& states);
 
     ggml_backend_buffer_type_t params_buffer_type_for(const TensorState& state) const;
+    ggml_backend_buffer_type_t split_buffer_type_for(const TensorState& state) const;
     void release_compute_staging_blocks(bool force                                            = false,
                                         const std::unordered_set<TensorState*>* target_states = nullptr);
     void release_params_storage_blocks(bool force                                            = false,
@@ -114,6 +118,9 @@ public:
     void set_writable_mmap(bool writable_mmap) { writable_mmap_ = writable_mmap; }
     void set_common_ignore_tensors(std::set<std::string> ignore_tensors);
     void set_loras(std::vector<LoraSpec> loras, SDVersion version);
+    void set_split_buffer_type(ggml_backend_t compute_backend, ggml_backend_buffer_type_t split_buft);
+
+    static bool tensor_shape_supports_split_buffer(const ggml_tensor* tensor);
 
     std::set<std::string> tensor_names() const;
 
@@ -122,7 +129,8 @@ public:
                                 ResidencyMode residency_mode,
                                 ggml_backend_t compute_backend,
                                 ggml_backend_t params_backend,
-                                size_t* registered_tensor_size = nullptr);
+                                size_t* registered_tensor_size = nullptr,
+                                bool allow_split_buffer        = false);
 
     template <typename Runner>
     bool register_runner_params(const std::string& desc,

src/runtime/denoiser.hpp

@@ -4,7 +4,10 @@
 #include <algorithm>
 #include <cctype>
 #include <cmath>
+#include <cstring>
 #include <functional>
+#include <limits>
+#include <map>
 #include <string>
 #include <utility>
 
@@ -302,6 +305,137 @@ struct KarrasScheduler : SigmaScheduler {
     }
 };
 
+struct BetaScheduler : SigmaScheduler {
+    static constexpr double alpha = 0.6;
+    static constexpr double beta  = 0.6;
+
+    static double log_beta(double a, double b) {
+        return std::lgamma(a) + std::lgamma(b) - std::lgamma(a + b);
+    }
+
+    static double incbeta(double x, double a, double b) {
+        if (x <= 0.0) {
+            return 0.0;
+        }
+        if (x >= 1.0) {
+            return 1.0;
+        }
+
+        // Continued fraction approximation using Lentz's method.
+        const int max_iter   = 200;
+        const double epsilon = 3.0e-7;
+        const double tiny    = 1e-30;
+
+        const double qab = a + b;
+        const double qap = a + 1.0;
+        const double qam = a - 1.0;
+
+        double c = 1.0;
+        double d = 1.0 - qab * x / qap;
+        if (std::abs(d) < tiny) {
+            d = tiny;
+        }
+        d        = 1.0 / d;
+        double h = d;
+
+        for (int m = 1; m <= max_iter; m++) {
+            const int m2 = 2 * m;
+
+            double aa = m * (b - m) * x / ((qam + m2) * (a + m2));
+            d         = 1.0 + aa * d;
+            if (std::abs(d) < tiny) {
+                d = tiny;
+            }
+            c = 1.0 + aa / c;
+            if (std::abs(c) < tiny) {
+                c = tiny;
+            }
+            d = 1.0 / d;
+            h *= d * c;
+
+            aa = -(a + m) * (qab + m) * x / ((a + m2) * (qap + m2));
+            d  = 1.0 + aa * d;
+            if (std::abs(d) < tiny) {
+                d = tiny;
+            }
+            c = 1.0 + aa / c;
+            if (std::abs(c) < tiny) {
+                c = tiny;
+            }
+            d                = 1.0 / d;
+            const double del = d * c;
+            h *= del;
+
+            if (std::abs(del - 1.0) < epsilon) {
+                break;
+            }
+        }
+
+        return std::exp(a * std::log(x) + b * std::log(1.0 - x) - log_beta(a, b)) / a * h;
+    }
+
+    static double beta_cdf(double x, double a, double b) {
+        if (x == 0.0) {
+            return 0.0;
+        }
+        if (x == 1.0) {
+            return 1.0;
+        }
+        if (x < (a + 1.0) / (a + b + 2.0)) {
+            return incbeta(x, a, b);
+        }
+        return 1.0 - incbeta(1.0 - x, b, a);
+    }
+
+    static double beta_ppf(double u, double a, double b, int max_iter = 30) {
+        double x = 0.5;
+        for (int i = 0; i < max_iter; i++) {
+            const double f = beta_cdf(x, a, b) - u;
+            if (std::abs(f) < 1e-10) {
+                break;
+            }
+            const double df = std::exp((a - 1.0) * std::log(x) + (b - 1.0) * std::log(1.0 - x) - log_beta(a, b));
+            x -= f / df;
+            if (x <= 0.0) {
+                x = 1e-10;
+            }
+            if (x >= 1.0) {
+                x = 1.0 - 1e-10;
+            }
+        }
+        return x;
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, float /*sigma_min*/, float /*sigma_max*/, t_to_sigma_t t_to_sigma) override {
+        std::vector<float> result;
+        result.reserve(n + 1);
+
+        const int t_max = TIMESTEPS - 1;
+        if (n == 0) {
+            return result;
+        } else if (n == 1) {
+            result.push_back(t_to_sigma(static_cast<float>(t_max)));
+            result.push_back(0.f);
+            return result;
+        }
+
+        int last_t = -1;
+        for (uint32_t i = 0; i < n; i++) {
+            const double u      = 1.0 - static_cast<double>(i) / static_cast<double>(n);
+            const double t_cont = beta_ppf(u, alpha, beta) * t_max;
+            const int t         = static_cast<int>(std::lround(t_cont));
+
+            if (t != last_t) {
+                result.push_back(t_to_sigma(static_cast<float>(t)));
+                last_t = t;
+            }
+        }
+
+        result.push_back(0.f);
+        return result;
+    }
+};
+
 struct SimpleScheduler : SigmaScheduler {
     std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         std::vector<float> result_sigmas;
@@ -882,6 +1016,7 @@ struct Denoiser {
                                             const sd::Tensor<float>& latent)         = 0;
     virtual sd::Tensor<float> inverse_noise_scaling(float sigma,
                                                     const sd::Tensor<float>& latent) = 0;
+    virtual float noise_level_to_sigma(float noise_level)                            = 0;
 
     virtual std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler_type, SDVersion version, const char* extra_sample_args = nullptr) {
         auto bound_t_to_sigma = std::bind(&Denoiser::t_to_sigma, this, std::placeholders::_1);
@@ -895,6 +1030,10 @@ struct Denoiser {
                 LOG_INFO("get_sigmas with Karras scheduler");
                 scheduler = std::make_shared<KarrasScheduler>();
                 break;
+            case BETA_SCHEDULER:
+                LOG_INFO("get_sigmas with Beta scheduler");
+                scheduler = std::make_shared<BetaScheduler>();
+                break;
             case EXPONENTIAL_SCHEDULER:
                 LOG_INFO("get_sigmas exponential scheduler");
                 scheduler = std::make_shared<ExponentialScheduler>();
@@ -1025,6 +1164,10 @@ struct CompVisDenoiser : public Denoiser {
         SD_UNUSED(sigma);
         return latent;
     }
+
+    float noise_level_to_sigma(float noise_level) {
+        return noise_level / (1.0f - noise_level);
+    }
 };
 
 struct CompVisVDenoiser : public CompVisDenoiser {
@@ -1112,6 +1255,10 @@ struct DiscreteFlowDenoiser : public Denoiser {
     sd::Tensor<float> inverse_noise_scaling(float sigma, const sd::Tensor<float>& latent) override {
         return latent * (1.0f / (1.0f - sigma));
     }
+
+    float noise_level_to_sigma(float noise_level) {
+        return noise_level;
+    }
 };
 
 struct FluxFlowDenoiser : public DiscreteFlowDenoiser {
@@ -1203,6 +1350,73 @@ struct SefiFlowDenoiser : public FluxFlowDenoiser {
     }
 };
 
+// MiniT2I predicts x0 directly and integrates a linear flow ODE:
+//   x_{t+dt} = x_t + (x0 - x_t)/(1 - t) * dt,  t in [0, 1), x0 = start = noise * 2.
+// Mapping sigma = 1 - t makes the generic Euler update
+//   x += (x - denoised)/sigma * (sigma_next - sigma)
+// exactly reproduce that step when denoised == x0. To make the generic
+// `denoised = pred * c_out + x * c_skip` yield x0 from the model's raw x0
+// prediction we use c_skip = 0, c_out = 1, c_in = 1. Sigmas run linearly 1 -> 0.
+struct MiniT2IFlowDenoiser : public Denoiser {
+    float sigma_min() override {
+        return 0.0f;
+    }
+
+    float sigma_max() override {
+        return 1.0f;
+    }
+
+    float sigma_to_t(float sigma) override {
+        return 1.0f - sigma;
+    }
+
+    float t_to_sigma(float t) override {
+        return 1.0f - t;
+    }
+
+    std::vector<float> get_scalings(float sigma) override {
+        SD_UNUSED(sigma);
+        float c_skip = 0.0f;
+        float c_out  = 1.0f;
+        float c_in   = 1.0f;
+        return {c_skip, c_out, c_in};
+    }
+
+    sd::Tensor<float> noise_scaling(float sigma,
+                                    const sd::Tensor<float>& noise,
+                                    const sd::Tensor<float>& latent) override {
+        SD_UNUSED(sigma);
+        SD_UNUSED(latent);
+        // Sampling starts from x0_init = noise * 2 (see MiniT2I reference).
+        return noise * 2.0f;
+    }
+
+    sd::Tensor<float> inverse_noise_scaling(float sigma, const sd::Tensor<float>& latent) override {
+        SD_UNUSED(sigma);
+        return latent;
+    }
+
+    float noise_level_to_sigma(float noise_level) {
+        SD_UNUSED(noise_level);
+        return 1.0f;
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler_type, SDVersion version, const char* extra_sample_args = nullptr) override {
+        SD_UNUSED(image_seq_len);
+        SD_UNUSED(scheduler_type);
+        SD_UNUSED(version);
+        SD_UNUSED(extra_sample_args);
+        // Uniform t schedule 0 -> 1 => sigma 1 -> 0, matching the reference loop.
+        std::vector<float> sigmas;
+        sigmas.reserve(n + 1);
+        for (uint32_t i = 0; i < n; ++i) {
+            sigmas.push_back(1.0f - static_cast<float>(i) / static_cast<float>(n));
+        }
+        sigmas.push_back(0.0f);
+        return sigmas;
+    }
+};
+
 typedef std::function<sd::guidance::GuiderOutput(const sd::Tensor<float>&, float, int)> denoise_cb_t;
 
 static std::pair<float, float> get_ancestral_step(float sigma_from,
@@ -1617,6 +1831,204 @@ static sd::Tensor<float> sample_dpmpp_2m_v2(denoise_cb_t model,
     return x;
 }
 
+// DPM-Solver++(2M) SDE, midpoint variant. Ref: Lu et al. arXiv:2211.01095;
+// k-diffusion sample_dpmpp_2m_sde.
+static sd::Tensor<float> sample_dpmpp_2m_sde(denoise_cb_t model,
+                                             sd::Tensor<float> x,
+                                             const std::vector<float>& sigmas,
+                                             std::shared_ptr<RNG> rng,
+                                             float eta) {
+    sd::Tensor<float> old_denoised;
+    bool have_old_denoised = false;
+    float h_last           = 0.f;
+
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        auto denoised_opt = model(x, sigmas[i], i + 1);
+        if (denoised_opt.pred.empty()) {
+            return {};
+        }
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
+
+        if (sigmas[i + 1] == 0.f) {
+            x = denoised;
+        } else {
+            float t     = -std::log(sigmas[i]);
+            float s     = -std::log(sigmas[i + 1]);
+            float h     = s - t;
+            float eta_h = eta * h;
+            float a     = sigmas[i + 1] / sigmas[i] * std::exp(-eta_h);
+            float b     = -std::expm1(-h - eta_h);
+
+            x = a * x + b * denoised;
+
+            if (have_old_denoised) {
+                float r = h_last / h;
+                x += (0.5f * b / r) * (denoised - old_denoised);
+            }
+            if (eta > 0.f) {
+                x += sd::Tensor<float>::randn_like(x, rng) * (sigmas[i + 1] * std::sqrt(-std::expm1(-2.f * eta_h)));
+            }
+            h_last = h;
+        }
+        old_denoised      = denoised;
+        have_old_denoised = true;
+    }
+    return x;
+}
+
+// Seeded Brownian tree providing deterministic, step-count-stable Gaussian
+// increments for stochastic samplers. Constructed once per generation; each
+// call returns unit-variance noise for interval [sigma_a, sigma_b].
+// Reference: torchsde BrownianTree; k-diffusion BatchedBrownianTree.
+class BrownianTreeNoiseSampler {
+public:
+    BrownianTreeNoiseSampler(const sd::Tensor<float>& x_template,
+                             double sigma_min,
+                             double sigma_max,
+                             uint64_t seed)
+        : t_min_(sigma_min),
+          t_max_(sigma_max),
+          shape_(x_template.shape()),
+          root_seed_(mix64(seed, 0x9E3779B97F4A7C15ULL)) {
+        auto rng = std::make_shared<STDDefaultRNG>();
+        rng->manual_seed(mix64(seed, 0xBF58476D1CE4E5B9ULL));
+        w_at_tmax_ = sd::Tensor<float>::randn(shape_, rng) * std::sqrt(static_cast<float>(t_max_ - t_min_));
+    }
+
+    sd::Tensor<float> operator()(double sigma_a, double sigma_b) {
+        double a          = clamp(std::min(sigma_a, sigma_b));
+        double b          = clamp(std::max(sigma_a, sigma_b));
+        auto dW           = w(b) - w(a);
+        float span        = static_cast<float>(std::max(std::abs(sigma_b - sigma_a), 1e-12));
+        return dW * (1.0f / std::sqrt(span));
+    }
+
+private:
+    static constexpr int kMaxDepth = 24;
+
+    static uint64_t mix64(uint64_t v, uint64_t salt) {
+        uint64_t z = v + salt;
+        z          = (z ^ (z >> 30)) * 0xBF58476D1CE4E5B9ULL;
+        z          = (z ^ (z >> 27)) * 0x94D049BB133111EBULL;
+        return z ^ (z >> 31);
+    }
+
+    double clamp(double t) const {
+        return std::min(std::max(t, t_min_), t_max_);
+    }
+
+    sd::Tensor<float> w(double t) {
+        auto it = cache_.find(t);
+        if (it != cache_.end()) {
+            return it->second;
+        }
+        sd::Tensor<float> zero = sd::Tensor<float>::zeros(shape_);
+        sd::Tensor<float> out  = bridge(t_min_, t_max_, zero, w_at_tmax_, t, root_seed_, kMaxDepth);
+        cache_.emplace(t, out);
+        return out;
+    }
+
+    sd::Tensor<float> bridge(double a,
+                             double c,
+                             const sd::Tensor<float>& w_a,
+                             const sd::Tensor<float>& w_c,
+                             double t,
+                             uint64_t node_seed,
+                             int depth) {
+        if (depth <= 0 || c - a < 1e-9) {
+            float alpha = (c > a) ? static_cast<float>((t - a) / (c - a)) : 0.5f;
+            return (1.0f - alpha) * w_a + alpha * w_c;
+        }
+        double m       = 0.5 * (a + c);
+        double std_dev = std::sqrt((c - m) * (m - a) / (c - a));
+        auto rng       = std::make_shared<STDDefaultRNG>();
+        rng->manual_seed(node_seed);
+        auto z   = sd::Tensor<float>::randn(shape_, rng);
+        auto w_m = 0.5f * (w_a + w_c) + static_cast<float>(std_dev) * z;
+        if (t == m) {
+            return w_m;
+        }
+        if (t < m) {
+            return bridge(a, m, w_a, w_m, t, mix64(node_seed, 1), depth - 1);
+        }
+        return bridge(m, c, w_m, w_c, t, mix64(node_seed, 2), depth - 1);
+    }
+
+    double t_min_;
+    double t_max_;
+    std::vector<int64_t> shape_;
+    uint64_t root_seed_;
+    sd::Tensor<float> w_at_tmax_;
+    std::map<double, sd::Tensor<float>> cache_;
+};
+
+// DPM-Solver++(2M) SDE, midpoint variant, with step-count-stable Brownian-tree
+// noise. Same trajectory shape at any step count for a given seed. Aliased in
+// k-diffusion / ComfyUI as sample_dpmpp_2m_sde_gpu.
+// Ref: Lu et al. arXiv:2211.01095; torchsde BrownianTree.
+static sd::Tensor<float> sample_dpmpp_2m_sde_bt(denoise_cb_t model,
+                                                sd::Tensor<float> x,
+                                                const std::vector<float>& sigmas,
+                                                std::shared_ptr<RNG> rng,
+                                                float eta) {
+    double sigma_max = 0.0;
+    double sigma_min = std::numeric_limits<double>::infinity();
+    for (float s : sigmas) {
+        if (s > 0.0f) {
+            sigma_max = std::max(sigma_max, static_cast<double>(s));
+            sigma_min = std::min(sigma_min, static_cast<double>(s));
+        }
+    }
+    if (sigma_max <= sigma_min) {
+        return x;
+    }
+    uint64_t tree_seed = 0;
+    {
+        auto draw = rng->randn(2);
+        std::memcpy(&tree_seed, draw.data(), sizeof(tree_seed));
+    }
+    BrownianTreeNoiseSampler noise_sampler(x, sigma_min, sigma_max, tree_seed);
+
+    sd::Tensor<float> old_denoised;
+    bool have_old_denoised = false;
+    float h_last           = 0.f;
+
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        auto denoised_opt = model(x, sigmas[i], i + 1);
+        if (denoised_opt.pred.empty()) {
+            return {};
+        }
+        sd::Tensor<float> denoised = std::move(denoised_opt.pred);
+
+        if (sigmas[i + 1] == 0.f) {
+            x = denoised;
+        } else {
+            float t     = -std::log(sigmas[i]);
+            float s     = -std::log(sigmas[i + 1]);
+            float h     = s - t;
+            float eta_h = eta * h;
+            float a     = sigmas[i + 1] / sigmas[i] * std::exp(-eta_h);
+            float b     = -std::expm1(-h - eta_h);
+
+            x = a * x + b * denoised;
+
+            if (have_old_denoised) {
+                float r = h_last / h;
+                x += (0.5f * b / r) * (denoised - old_denoised);
+            }
+            if (eta > 0.f) {
+                x += noise_sampler(sigmas[i], sigmas[i + 1]) * (sigmas[i + 1] * std::sqrt(-std::expm1(-2.f * eta_h)));
+            }
+            h_last = h;
+        }
+        old_denoised      = denoised;
+        have_old_denoised = true;
+    }
+    return x;
+}
+
 using SamplerExtraArgs = KeyValueArgs;
 
 static sd::Tensor<float> sample_lcm(denoise_cb_t model,
@@ -2293,6 +2705,10 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
             return sample_res_2s(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case ER_SDE_SAMPLE_METHOD:
             return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
+        case DPMPP2M_SDE_SAMPLE_METHOD:
+            return sample_dpmpp_2m_sde(model, std::move(x), sigmas, rng, eta);
+        case DPMPP2M_SDE_BT_SAMPLE_METHOD:
+            return sample_dpmpp_2m_sde_bt(model, std::move(x), sigmas, rng, eta);
         case DDIM_TRAILING_SAMPLE_METHOD:
             // DDIM is equivalent to Euler Ancestral with the Simple scheduler
             return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);

src/runtime/imatrix.cpp

@@ -0,0 +1,308 @@
+#include "runtime/imatrix.h"
+
+/* Adapted from llama.cpp (credits: Kawrakow). */
+
+#include "core/util.h"
+#include "ggml-backend.h"
+#include "ggml.h"
+#include "stable-diffusion.h"
+
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+
+static IMatrixCollector imatrix_collector;
+
+IMatrixCollector& get_imatrix_collector() {
+    return imatrix_collector;
+}
+
+// remove any prefix and suffixes from the name
+// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+static std::string filter_tensor_name(const char* name) {
+    std::string wname;
+    const char* p = strchr(name, '#');
+    if (p != NULL) {
+        p             = p + 1;
+        const char* q = strchr(p, '#');
+        if (q != NULL) {
+            wname = std::string(p, q - p);
+        } else {
+            wname = p;
+        }
+    } else {
+        wname = name;
+    }
+    return wname;
+}
+
+bool IMatrixCollector::collect_imatrix(struct ggml_tensor* t, bool ask, void* user_data) {
+    GGML_UNUSED(user_data);
+    if (t == nullptr) {
+        return false;
+    }
+    if (t->op != GGML_OP_MUL_MAT && t->op != GGML_OP_MUL_MAT_ID) {
+        return false;
+    }
+
+    const struct ggml_tensor* src0 = t->src[0];
+    const struct ggml_tensor* src1 = t->src[1];
+    if (src0 == nullptr || src1 == nullptr) {
+        return false;
+    }
+    std::string wname = filter_tensor_name(src0->name);
+
+    // when ask is true, the scheduler wants to know if we are interested in data from this tensor
+    // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
+    if (ask) {
+        if (t->op == GGML_OP_MUL_MAT_ID) {
+            return true;  // collect all indirect matrix multiplications
+        }
+        // why are small batches ignored (<16 tokens)?
+        // if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
+        if (!(wname.substr(0, 6) == "model." || wname.substr(0, 17) == "cond_stage_model." || wname.substr(0, 14) == "text_encoders.")) {
+            return false;
+        }
+        return true;
+    }
+    std::lock_guard<std::mutex> lock(mutex_);
+
+    // copy the data from the GPU memory if needed
+    const bool is_host = src1->buffer == NULL || ggml_backend_buffer_is_host(src1->buffer);
+
+    if (!is_host) {
+        src1_data_.resize(ggml_nelements(src1));
+        ggml_backend_tensor_get(src1, src1_data_.data(), 0, ggml_nbytes(src1));
+    }
+
+    const float* data = is_host ? (const float*)src1->data : src1_data_.data();
+
+    // this has been adapted to the new format of storing merged experts in a single 3d tensor
+    // ref: https://github.com/ggml-org/llama.cpp/pull/6387
+    if (t->op == GGML_OP_MUL_MAT_ID) {
+        //   ids  -> [n_experts_used, n_tokens]
+        //   src1 -> [cols, n_expert_used, n_tokens]
+        const ggml_tensor* ids = t->src[2];
+        const int n_as         = static_cast<int>(src0->ne[2]);
+        const int n_ids        = static_cast<int>(ids->ne[0]);
+
+        // the top-k selected expert ids are stored in the ids tensor
+        // for simplicity, always copy ids to host, because it is small
+        // take into account that ids is not contiguous!
+
+        GGML_ASSERT(ids->ne[1] == src1->ne[2]);
+
+        ids_.resize(ggml_nbytes(ids));
+        ggml_backend_tensor_get(ids, ids_.data(), 0, ggml_nbytes(ids));
+
+        auto& e = stats_[wname];
+
+        ++e.ncall;
+
+        if (e.values.empty()) {
+            e.values.resize(src1->ne[0] * n_as, 0);
+            e.counts.resize(src1->ne[0] * n_as, 0);
+        } else if (e.values.size() != (size_t)src1->ne[0] * n_as) {
+            LOG_ERROR("inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0] * n_as);
+            exit(1);  // GGML_ABORT("fatal error");
+        }
+        // loop over all possible experts, regardless if they are used or not in the batch
+        for (int ex = 0; ex < n_as; ++ex) {
+            size_t e_start = ex * src1->ne[0];
+
+            for (int idx = 0; idx < n_ids; ++idx) {
+                for (int row = 0; row < (int)src1->ne[2]; ++row) {
+                    const int excur = *(const int32_t*)(ids_.data() + row * ids->nb[1] + idx * ids->nb[0]);
+
+                    GGML_ASSERT(excur >= 0 && excur < n_as);  // sanity check
+
+                    if (excur != ex)
+                        continue;
+
+                    const int64_t i11 = idx % src1->ne[1];
+                    const int64_t i12 = row;
+                    const float* x    = (const float*)((const char*)data + i11 * src1->nb[1] + i12 * src1->nb[2]);
+
+                    for (int j = 0; j < (int)src1->ne[0]; ++j) {
+                        e.values[e_start + j] += x[j] * x[j];
+                        e.counts[e_start + j]++;
+                        if (!std::isfinite(e.values[e_start + j])) {
+                            LOG_ERROR("%f detected in %s\n", e.values[e_start + j], wname.c_str());
+                            exit(1);
+                        }
+                    }
+                }
+            }
+        }
+    } else {
+        auto& e = stats_[wname];
+        if (e.values.empty()) {
+            e.values.resize(src1->ne[0], 0);
+            e.counts.resize(src1->ne[0], 0);
+        } else if (e.values.size() != (size_t)src1->ne[0]) {
+            LOG_WARN("inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
+            exit(1);  // GGML_ABORT("fatal error");
+        }
+
+        ++e.ncall;
+        for (int row = 0; row < (int)src1->ne[1]; ++row) {
+            const float* x = data + row * src1->ne[0];
+            for (int j = 0; j < (int)src1->ne[0]; ++j) {
+                if (std::isfinite(x[j])) {
+                    e.values[j] += x[j] * x[j];
+                    e.counts[j]++;
+                    if (!std::isfinite(e.values[j])) {
+                        LOG_WARN("%f detected in %s\n", e.values[j], wname.c_str());
+                        exit(1);
+                    }
+                } else {
+                    // Likely something from an attention mask?
+                }
+            }
+        }
+    }
+    return true;
+}
+
+bool load_imatrix(const char* imatrix_path) {
+    return imatrix_collector.load_imatrix(imatrix_path);
+}
+
+void save_imatrix(const char* imatrix_path) {
+    imatrix_collector.save_imatrix(imatrix_path);
+}
+
+static bool collect_imatrix(struct ggml_tensor* t, bool ask, void* user_data) {
+    return imatrix_collector.collect_imatrix(t, ask, user_data);
+}
+
+void enable_imatrix_collection() {
+    sd_set_backend_eval_callback(collect_imatrix, nullptr);
+}
+
+void disable_imatrix_collection() {
+    sd_set_backend_eval_callback(nullptr, nullptr);
+}
+
+void IMatrixCollector::save_imatrix(std::string fname, int ncall) const {
+    if (ncall > 0) {
+        fname += ".at_";
+        fname += std::to_string(ncall);
+    }
+    // avoid writing imatrix entries that do not have full data
+    // this can happen with MoE models where some of the experts end up not being exercised by the provided training data
+
+    int n_entries = 0;
+    std::vector<std::string> to_store;
+
+    for (const auto& kv : stats_) {
+        const int n_all = static_cast<int>(kv.second.counts.size());
+
+        if (n_all == 0) {
+            continue;
+        }
+
+        int n_zeros = 0;
+        for (const int c : kv.second.counts) {
+            if (c == 0) {
+                n_zeros++;
+            }
+        }
+
+        if (n_zeros == n_all) {
+            LOG_WARN("entry '%40s' has no data - skipping\n", kv.first.c_str());
+            continue;
+        }
+
+        if (n_zeros > 0) {
+            LOG_WARN("entry '%40s' has partial data (%.2f%%) - skipping\n", kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
+            continue;
+        }
+
+        n_entries++;
+        to_store.push_back(kv.first);
+    }
+
+    if (to_store.size() < stats_.size()) {
+        LOG_WARN("storing only %zu out of %zu entries\n", to_store.size(), stats_.size());
+    }
+
+    std::ofstream out(fname, std::ios::binary);
+    out.write((const char*)&n_entries, sizeof(n_entries));
+    for (const auto& name : to_store) {
+        const auto& stat = stats_.at(name);
+        int len          = static_cast<int>(name.size());
+        out.write((const char*)&len, sizeof(len));
+        out.write(name.c_str(), len);
+        out.write((const char*)&stat.ncall, sizeof(stat.ncall));
+        int nval = static_cast<int>(stat.values.size());
+        out.write((const char*)&nval, sizeof(nval));
+        if (nval > 0) {
+            std::vector<float> tmp(nval);
+            for (int i = 0; i < nval; i++) {
+                tmp[i] = (stat.values[i] / static_cast<float>(stat.counts[i])) * static_cast<float>(stat.ncall);
+            }
+            out.write((const char*)tmp.data(), nval * sizeof(float));
+        }
+    }
+
+    // Write the number of call the matrix was computed with
+    out.write((const char*)&last_call_, sizeof(last_call_));
+}
+
+bool IMatrixCollector::load_imatrix(const char* fname) {
+    std::ifstream in(fname, std::ios::binary);
+    if (!in) {
+        LOG_ERROR("failed to open %s\n", fname);
+        return false;
+    }
+    int n_entries;
+    in.read((char*)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        LOG_ERROR("no data in file %s\n", fname);
+        return false;
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len;
+        in.read((char*)&len, sizeof(len));
+        std::vector<char> name_as_vec(len + 1);
+        in.read((char*)name_as_vec.data(), len);
+        if (in.fail()) {
+            LOG_ERROR("failed reading name for entry %d from %s\n", i + 1, fname);
+            return false;
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto& e = stats_[std::move(name)];
+        int ncall;
+        in.read((char*)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char*)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            LOG_ERROR("failed reading number of values for entry %d\n", i);
+            stats_ = {};
+            return false;
+        }
+
+        if (e.values.empty()) {
+            e.values.resize(nval, 0);
+            e.counts.resize(nval, 0);
+        }
+
+        std::vector<float> tmp(nval);
+        in.read((char*)tmp.data(), nval * sizeof(float));
+        if (in.fail()) {
+            LOG_ERROR("failed reading data for entry %d\n", i);
+            stats_ = {};
+            return false;
+        }
+
+        // Recreate the state as expected by save_imatrix(), and correct for weighted sum.
+        for (int i = 0; i < nval; i++) {
+            e.values[i] += tmp[i];
+            e.counts[i] += ncall;
+        }
+        e.ncall += ncall;
+    }
+    return true;
+}

src/runtime/imatrix.h

@@ -0,0 +1,45 @@
+#ifndef __SD_RUNTIME_IMATRIX_H__
+#define __SD_RUNTIME_IMATRIX_H__
+
+#include <fstream>
+#include <mutex>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+/* Adapted from llama.cpp (credits: Kawrakow). */
+
+struct ggml_tensor;
+
+struct IMatrixStats {
+    std::vector<float> values{};
+    std::vector<int> counts{};
+    int ncall = 0;
+};
+
+class IMatrixCollector {
+private:
+    std::unordered_map<std::string, IMatrixStats> stats_ = {};
+    std::mutex mutex_;
+    int last_call_ = 0;
+    std::vector<float> src1_data_;
+    std::vector<char> ids_;  // the expert ids from ggml_mul_mat_id
+
+public:
+    IMatrixCollector() = default;
+    bool collect_imatrix(struct ggml_tensor* t, bool ask, void* user_data);
+    void save_imatrix(std::string fname, int ncall = -1) const;
+    bool load_imatrix(const char* fname);
+    std::vector<float> get_values(const std::string& key) const {
+        auto it = stats_.find(key);
+        if (it != stats_.end()) {
+            return it->second.values;
+        } else {
+            return {};
+        }
+    }
+};
+
+IMatrixCollector& get_imatrix_collector();
+
+#endif  // __SD_RUNTIME_IMATRIX_H__

src/upscaler.cpp

@@ -4,6 +4,7 @@
 #include "model_loader.h"
 #include "stable-diffusion.h"
 
+#include <cstdlib>
 #include <utility>
 
 UpscalerGGML::UpscalerGGML(int n_threads,
@@ -45,6 +46,7 @@ bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
     std::string error;
     if (!backend_manager.init(backend_spec.c_str(),
                               params_backend_spec.c_str(),
+                              /*split_mode_spec=*/nullptr,
                               &error)) {
         LOG_ERROR("upscaler backend config failed: %s", error.c_str());
         return false;
@@ -198,8 +200,41 @@ upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
     return upscaler_ctx;
 }
 
-sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor) {
-    return upscaler_ctx->upscaler->upscale(input_image, upscale_factor);
+bool upscale(upscaler_ctx_t* upscaler_ctx,
+             sd_image_t input_image,
+             uint32_t upscale_factor,
+             sd_image_t** images_out,
+             int* num_images_out) {
+    if (images_out != nullptr) {
+        *images_out = nullptr;
+    }
+    if (num_images_out != nullptr) {
+        *num_images_out = 0;
+    }
+    if (upscaler_ctx == nullptr || upscaler_ctx->upscaler == nullptr) {
+        return false;
+    }
+
+    sd_image_t* result_images = (sd_image_t*)calloc(1, sizeof(sd_image_t));
+    if (result_images == nullptr) {
+        return false;
+    }
+
+    result_images[0] = upscaler_ctx->upscaler->upscale(input_image, upscale_factor);
+    if (result_images[0].data == nullptr) {
+        free(result_images);
+        return false;
+    }
+
+    if (num_images_out != nullptr) {
+        *num_images_out = 1;
+    }
+    if (images_out != nullptr) {
+        *images_out = result_images;
+    } else {
+        free_sd_images(result_images, 1);
+    }
+    return true;
 }
 
 int get_upscale_factor(upscaler_ctx_t* upscaler_ctx) {