refactor: simplify the logic of pm id image loading (#827)

* opt tensor loading

* fix build failure

* revert the changes

* allow the use of n_threads

* fix lora loading

* optimize lora loading

* add mutex

* use atomic

* fix build

* fix potential duplicate issue

* avoid duplicate lookup of lora tensor

* fix progeress bar

* remove unused remove_duplicates

---------

Co-authored-by: leejet <leejet714@gmail.com>

.github/workflows/build.yml

@@ -163,7 +163,7 @@ jobs:
           - build: "avx512"
             defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
           - build: "cuda12"
-            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;80;75"
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;86;80;75"
           # - build: "rocm5.5"
           #   defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
           - build: 'vulkan'

.gitignore

@@ -1,13 +1,14 @@
 build*/
+cmake-build-*/
 test/
 .vscode/
+.idea/
 .cache/
 *.swp
-.vscode/
 *.bat
 *.bin
 *.exe
 *.gguf
 output*.png
 models*
-*.log
+*.log

.gitmodules

@@ -1,3 +1,3 @@
 [submodule "ggml"]
     path = ggml
-	url = https://github.com/ggerganov/ggml.git
+	url = https://github.com/ggml-org/ggml.git

CMakeLists.txt

@@ -33,6 +33,7 @@ option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_MUSA                       "sd: musa backend" OFF)
 option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
+option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
 if(SD_CUDA)
@@ -118,23 +119,37 @@ endif()
 
 set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
-# see https://github.com/ggerganov/ggml/pull/682
-add_definitions(-DGGML_MAX_NAME=128)
+if (NOT SD_USE_SYSTEM_GGML)
+    # see https://github.com/ggerganov/ggml/pull/682
+    add_definitions(-DGGML_MAX_NAME=128)
+endif()
 
 # deps
 # Only add ggml if it hasn't been added yet
 if (NOT TARGET ggml)
-    add_subdirectory(ggml)
+    if (SD_USE_SYSTEM_GGML)
+        find_package(ggml REQUIRED)
+        if (NOT ggml_FOUND)
+            message(FATAL_ERROR "System-installed GGML library not found.")
+        endif()
+        add_library(ggml ALIAS ggml::ggml)
+    else()
+        add_subdirectory(ggml)
+    endif()
 endif()
 
 add_subdirectory(thirdparty)
 
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
-target_compile_features(${SD_LIB} PUBLIC cxx_std_11)
+target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)
 
 
 if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
+set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)

Dockerfile.musa

@@ -1,6 +1,7 @@
-ARG MUSA_VERSION=rc3.1.1
+ARG MUSA_VERSION=rc4.2.0
+ARG UBUNTU_VERSION=22.04
 
-FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu22.04 as build
+FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build
 
 RUN apt-get update && apt-get install -y ccache cmake git
 
@@ -15,7 +16,7 @@ RUN mkdir build && cd build && \
         -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release && \
     cmake --build . --config Release
 
-FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu22.04 as runtime
+FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runtime
 
 COPY --from=build /sd.cpp/build/bin/sd /sd
 

Dockerfile.sycl

@@ -0,0 +1,19 @@
+ARG SYCL_VERSION=2025.1.0-0
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
+
+RUN apt-get update && apt-get install -y cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release -j$(nproc)
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

README.md

@@ -4,19 +4,33 @@
 
 # stable-diffusion.cpp
 
-Inference of Stable Diffusion and Flux in pure C/C++
+Diffusion model(SD,Flux,Wan,...) inference in pure C/C++
+
+***Note that this project is under active development. \
+API and command-line option may change frequently.***
 
 ## Features
 
 - Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
 - Super lightweight and without external dependencies
-- SD1.x, SD2.x, SDXL and [SD3/SD3.5](./docs/sd3.md) support
-    - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
-- [Flux-dev/Flux-schnell Support](./docs/flux.md)
-- [FLUX.1-Kontext-dev](./docs/kontext.md)
-- [Chroma](./docs/chroma.md)
-- [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo) and [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo) support
-- [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
+- Supported models
+  - Image Models
+    - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
+    - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
+      - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
+    - [SD3/SD3.5](./docs/sd3.md)
+    - [Flux-dev/Flux-schnell](./docs/flux.md)
+    - [Chroma](./docs/chroma.md)
+  - Image Edit Models
+    - [FLUX.1-Kontext-dev](./docs/kontext.md)
+  - Video Models
+    - [Wan2.1/Wan2.2](./docs/wan.md)
+  - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
+  - Control Net support with SD 1.5
+  - LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
+  - Latent Consistency Models support (LCM/LCM-LoRA)
+  - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
+  - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
 - 16-bit, 32-bit float support
 - 2-bit, 3-bit, 4-bit, 5-bit and 8-bit integer quantization support
 - Accelerated memory-efficient CPU inference
@@ -26,15 +40,9 @@ Inference of Stable Diffusion and Flux in pure C/C++
 - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
     - No need to convert to `.ggml` or `.gguf` anymore!
 - Flash Attention for memory usage optimization
-- Original `txt2img` and `img2img` mode
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
-- LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
-- Latent Consistency Models support (LCM/LCM-LoRA)
-- Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
-- Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
 - VAE tiling processing for reduce memory usage
-- Control Net support with SD 1.5
 - Sampling method
     - `Euler A`
     - `Euler`
@@ -52,14 +60,6 @@ Inference of Stable Diffusion and Flux in pure C/C++
     - Windows
     - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
 
-### TODO
-
-- [ ] More sampling methods
-- [ ] Make inference faster
-    - The current implementation of ggml_conv_2d is slow and has high memory usage
-- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
-- [ ] Implement Inpainting support
-
 ## Usage
 
 For most users, you can download the built executable program from the latest [release](https://github.com/leejet/stable-diffusion.cpp/releases/latest).
@@ -129,7 +129,9 @@ This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure
 Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
 
 ```
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
+export GFX_NAME=$(rocminfo | grep -m 1 -E "gfx[^0]{1}" | sed -e 's/ *Name: *//' | awk '{$1=$1; print}' || echo "rocminfo missing")
+echo $GFX_NAME
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
 cmake --build . --config Release
 ```
 
@@ -282,31 +284,36 @@ usage: ./bin/sd [arguments]
 
 arguments:
   -h, --help                         show this help message and exit
-  -M, --mode [MODE]                  run mode, one of: [img_gen, convert], default: img_gen
+  -M, --mode [MODE]                  run mode, one of: [img_gen, vid_gen, convert], default: img_gen
   -t, --threads N                    number of threads to use during computation (default: -1)
                                      If threads <= 0, then threads will be set to the number of CPU physical cores
+  --offload-to-cpu                   place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
   -m, --model [MODEL]                path to full model
   --diffusion-model                  path to the standalone diffusion model
+  --high-noise-diffusion-model       path to the standalone high noise diffusion model
   --clip_l                           path to the clip-l text encoder
   --clip_g                           path to the clip-g text encoder
+  --clip_vision                      path to the clip-vision encoder
   --t5xxl                            path to the t5xxl text encoder
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
   --control-net [CONTROL_PATH]       path to control net model
   --embd-dir [EMBEDDING_PATH]        path to embeddings
-  --stacked-id-embd-dir [DIR]        path to PHOTOMAKER stacked id embeddings
-  --input-id-images-dir [DIR]        path to PHOTOMAKER input id images dir
-  --normalize-input                  normalize PHOTOMAKER input id images
   --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now
   --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)
   --type [TYPE]                      weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K)
                                      If not specified, the default is the type of the weight file
   --tensor-type-rules [EXPRESSION]   weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --lora-model-dir [DIR]             lora model directory
-  -i, --init-img [IMAGE]             path to the input image, required by img2img
+  -i, --init-img [IMAGE]             path to the init image, required by img2img
   --mask [MASK]                      path to the mask image, required by img2img with mask
+  -i, --end-img [IMAGE]              path to the end image, required by flf2v
   --control-image [IMAGE]            path to image condition, control net
   -r, --ref-image [PATH]             reference image for Flux Kontext models (can be used multiple times)
+  --control-video [PATH]             path to control video frames, It must be a directory path.
+                                     The video frames inside should be stored as images in lexicographical (character) order
+                                     For example, if the control video path is `frames`, the directory contain images such as 00.png, 01.png, 鈥?etc.
+  --increase-ref-index               automatically increase the indices of references images based on the order they are listed (starting with 1).
   -o, --output OUTPUT                path to write result image to (default: ./output.png)
   -p, --prompt [PROMPT]              the prompt to render
   -n, --negative-prompt PROMPT       the negative prompt (default: "")
@@ -319,34 +326,64 @@ arguments:
   --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])
   --skip-layer-start START           SLG enabling point: (default: 0.01)
   --skip-layer-end END               SLG disabling point: (default: 0.2)
+  --scheduler {discrete, karras, exponential, ays, gits, smoothstep} Denoiser sigma scheduler (default: discrete)
+  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
+                                     sampling method (default: "euler" for Flux/SD3/Wan, "euler_a" otherwise)
+  --steps  STEPS                     number of sample steps (default: 20)
+  --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)
+  --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --high-noise-guidance SCALE        (high noise) distilled guidance scale for models with guidance input (default: 3.5)
+  --high-noise-slg-scale SCALE       (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium
+  --high-noise-eta SCALE             (high noise) eta in DDIM, only for DDIM and TCD: (default: 0)
+  --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])
+  --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)
+  --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)
+  --high-noise-scheduler {discrete, karras, exponential, ays, gits, smoothstep} Denoiser sigma scheduler (default: discrete)
+  --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
+                                     (high noise) sampling method (default: "euler_a")
+  --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)
                                      SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])
   --strength STRENGTH                strength for noising/unnoising (default: 0.75)
-  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20)
   --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
                                      1.0 corresponds to full destruction of information in init image
   -H, --height H                     image height, in pixel space (default: 512)
   -W, --width W                      image width, in pixel space (default: 512)
-  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
-                                     sampling method (default: "euler_a")
-  --steps  STEPS                     number of sample steps (default: 20)
   --rng {std_default, cuda}          RNG (default: cuda)
   -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
   -b, --batch-count COUNT            number of images to generate
-  --schedule {discrete, karras, exponential, ays, gits} Denoiser sigma schedule (default: discrete)
-  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
+  --clip-skip N                      ignore last_dot_pos layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
                                      <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   --vae-tiling                       process vae in tiles to reduce memory usage
+  --vae-tile-size [X]x[Y]            tile size for vae tiling (default: 32x32)
+  --vae-relative-tile-size [X]x[Y]   relative tile size for vae tiling, in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
+  --vae-tile-overlap OVERLAP         tile overlap for vae tiling, in fraction of tile size (default: 0.5)
   --vae-on-cpu                       keep vae in cpu (for low vram)
   --clip-on-cpu                      keep clip in cpu (for low vram)
   --diffusion-fa                     use flash attention in the diffusion model (for low vram)
                                      Might lower quality, since it implies converting k and v to f16.
                                      This might crash if it is not supported by the backend.
+  --diffusion-conv-direct            use Conv2d direct in the diffusion model
+                                     This might crash if it is not supported by the backend.
+  --vae-conv-direct                  use Conv2d direct in the vae model (should improve the performance)
+                                     This might crash if it is not supported by the backend.
   --control-net-cpu                  keep controlnet in cpu (for low vram)
   --canny                            apply canny preprocessor (edge detection)
   --color                            colors the logging tags according to level
   --chroma-disable-dit-mask          disable dit mask for chroma
   --chroma-enable-t5-mask            enable t5 mask for chroma
   --chroma-t5-mask-pad  PAD_SIZE     t5 mask pad size of chroma
+  --video-frames                     video frames (default: 1)
+  --fps                              fps (default: 24)
+  --moe-boundary BOUNDARY            timestep boundary for Wan2.2 MoE model. (default: 0.875)
+                                     only enabled if `--high-noise-steps` is set to -1
+  --flow-shift SHIFT                 shift value for Flow models like SD3.x or WAN (default: auto)
+  --vace-strength                    wan vace strength
+  --photo-maker                      path to PHOTOMAKER model
+  --pm-id-images-dir [DIR]           path to PHOTOMAKER input id images dir
+  --pm-id-embed-path [PATH]          path to PHOTOMAKER v2 id embed
+  --pm-style-strength                strength for keeping PHOTOMAKER input identity (default: 20)
+  --normalize-input                  normalize PHOTOMAKER input id images
   -v, --verbose                      print extra info
 ```
 
@@ -356,9 +393,9 @@ arguments:
 ./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
 # ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
 # ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
-# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v
-# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v
-# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v
+# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```
 
 Using formats of different precisions will yield results of varying quality.
@@ -434,3 +471,5 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 - [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)

clip.hpp

@@ -179,9 +179,9 @@ public:
 
         auto it = encoder.find(utf8_to_utf32("img</w>"));
         if (it != encoder.end()) {
-            LOG_DEBUG(" trigger word img already in vocab");
+            LOG_DEBUG("trigger word img already in vocab");
         } else {
-            LOG_DEBUG(" trigger word img not in vocab yet");
+            LOG_DEBUG("trigger word img not in vocab yet");
         }
 
         int rank = 0;
@@ -488,14 +488,14 @@ public:
         blocks["mlp"] = std::shared_ptr<GGMLBlock>(new CLIPMLP(d_model, intermediate_size));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, bool mask = true) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, ggml_backend_t backend, struct ggml_tensor* x, bool mask = true) {
         // x: [N, n_token, d_model]
         auto self_attn   = std::dynamic_pointer_cast<MultiheadAttention>(blocks["self_attn"]);
         auto layer_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm1"]);
         auto layer_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm2"]);
         auto mlp         = std::dynamic_pointer_cast<CLIPMLP>(blocks["mlp"]);
 
-        x = ggml_add(ctx, x, self_attn->forward(ctx, layer_norm1->forward(ctx, x), mask));
+        x = ggml_add(ctx, x, self_attn->forward(ctx, backend, layer_norm1->forward(ctx, x), mask));
         x = ggml_add(ctx, x, mlp->forward(ctx, layer_norm2->forward(ctx, x)));
         return x;
     }
@@ -517,7 +517,11 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, int clip_skip = -1, bool mask = true) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                int clip_skip = -1,
+                                bool mask     = true) {
         // x: [N, n_token, d_model]
         int layer_idx = n_layer - 1;
         // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -532,7 +536,7 @@ public:
             }
             std::string name = "layers." + std::to_string(i);
             auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
-            x                = layer->forward(ctx, x, mask);  // [N, n_token, d_model]
+            x                = layer->forward(ctx, backend, x, mask);  // [N, n_token, d_model]
             // LOG_DEBUG("layer %d", i);
         }
         return x;
@@ -544,10 +548,16 @@ protected:
     int64_t embed_dim;
     int64_t vocab_size;
     int64_t num_positions;
+    bool force_clip_f32;
 
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
-        enum ggml_type token_wtype    = GGML_TYPE_F32;  //(tensor_types.find(prefix + "token_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "token_embedding.weight"] : GGML_TYPE_F32;
-        enum ggml_type position_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type token_wtype = GGML_TYPE_F32;
+        if (!force_clip_f32) {
+            auto tensor_type = tensor_types.find(prefix + "token_embedding.weight");
+            if (tensor_type != tensor_types.end())
+                token_wtype = tensor_type->second;
+        }
+        enum ggml_type position_wtype = GGML_TYPE_F32;
 
         params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
         params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
@@ -556,10 +566,12 @@ protected:
 public:
     CLIPEmbeddings(int64_t embed_dim,
                    int64_t vocab_size    = 49408,
-                   int64_t num_positions = 77)
+                   int64_t num_positions = 77,
+                   bool force_clip_f32   = false)
         : embed_dim(embed_dim),
           vocab_size(vocab_size),
-          num_positions(num_positions) {
+          num_positions(num_positions),
+          force_clip_f32(force_clip_f32) {
     }
 
     struct ggml_tensor* get_token_embed_weight() {
@@ -594,10 +606,10 @@ protected:
     int64_t image_size;
     int64_t num_patches;
     int64_t num_positions;
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
-        enum ggml_type patch_wtype    = GGML_TYPE_F16;  // tensor_types.find(prefix + "patch_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "patch_embedding.weight"] : GGML_TYPE_F16;
-        enum ggml_type class_wtype    = GGML_TYPE_F32;  // tensor_types.find(prefix + "class_embedding") != tensor_types.end() ? tensor_types[prefix + "class_embedding"] : GGML_TYPE_F32;
-        enum ggml_type position_wtype = GGML_TYPE_F32;  // tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type patch_wtype    = GGML_TYPE_F16;
+        enum ggml_type class_wtype    = GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;
 
         params["patch_embedding.weight"]    = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
         params["class_embedding"]           = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
@@ -657,9 +669,9 @@ enum CLIPVersion {
 
 class CLIPTextModel : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
         if (version == OPEN_CLIP_VIT_BIGG_14) {
-            enum ggml_type wtype      = GGML_TYPE_F32;  // tensor_types.find(prefix + "text_projection") != tensor_types.end() ? tensor_types[prefix + "text_projection"] : GGML_TYPE_F32;
+            enum ggml_type wtype      = GGML_TYPE_F32;
             params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
         }
     }
@@ -674,12 +686,11 @@ public:
     int32_t n_head            = 12;
     int32_t n_layer           = 12;    // num_hidden_layers
     int32_t projection_dim    = 1280;  // only for OPEN_CLIP_VIT_BIGG_14
-    int32_t clip_skip         = -1;
     bool with_final_ln        = true;
 
     CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                   bool with_final_ln  = true,
-                  int clip_skip_value = -1)
+                  bool force_clip_f32 = false)
         : version(version), with_final_ln(with_final_ln) {
         if (version == OPEN_CLIP_VIT_H_14) {
             hidden_size       = 1024;
@@ -692,37 +703,31 @@ public:
             n_head            = 20;
             n_layer           = 32;
         }
-        set_clip_skip(clip_skip_value);
 
-        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token));
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token, force_clip_f32));
         blocks["encoder"]          = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
 
-    void set_clip_skip(int skip) {
-        if (skip <= 0) {
-            skip = -1;
-        }
-        clip_skip = skip;
-    }
-
     struct ggml_tensor* get_token_embed_weight() {
         auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         return embeddings->get_token_embed_weight();
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* tkn_embeddings,
                                 size_t max_token_idx = 0,
-                                bool return_pooled   = false) {
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
         // input_ids: [N, n_token]
         auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
         auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
 
         auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
-        x      = encoder->forward(ctx, x, return_pooled ? -1 : clip_skip, true);
+        x      = encoder->forward(ctx, backend, x, return_pooled ? -1 : clip_skip, true);
         if (return_pooled || with_final_ln) {
             x = final_layer_norm->forward(ctx, x);
         }
@@ -733,7 +738,7 @@ public:
             if (text_projection != NULL) {
                 pooled = ggml_nn_linear(ctx, pooled, text_projection, NULL);
             } else {
-                LOG_DEBUG("Missing text_projection matrix, assuming identity...");
+                LOG_DEBUG("identity projection");
             }
             return pooled;  // [hidden_size, 1, 1]
         }
@@ -774,7 +779,11 @@ public:
         blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values, bool return_pooled = true) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* pixel_values,
+                                bool return_pooled = true,
+                                int clip_skip      = -1) {
         // pixel_values: [N, num_channels, image_size, image_size]
         auto embeddings     = std::dynamic_pointer_cast<CLIPVisionEmbeddings>(blocks["embeddings"]);
         auto pre_layernorm  = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_layernorm"]);
@@ -783,7 +792,7 @@ public:
 
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
-        x      = encoder->forward(ctx, x, -1, false);
+        x      = encoder->forward(ctx, backend, x, clip_skip, false);
         // print_ggml_tensor(x, true, "ClipVisionModel x: ");
         auto last_hidden_state = x;
         x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
@@ -805,8 +814,8 @@ protected:
     int64_t out_features;
     bool transpose_weight;
 
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
-        enum ggml_type wtype = tensor_types.find(prefix + "weight") != tensor_types.end() ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
         if (transpose_weight) {
             params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
         } else {
@@ -851,16 +860,23 @@ public:
         blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* pixel_values,
+                                bool return_pooled = true,
+                                int clip_skip      = -1) {
         // pixel_values: [N, num_channels, image_size, image_size]
-        // return: [N, projection_dim]
+        // return: [N, projection_dim] if return_pooled else [N, n_token, hidden_size]
         auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
         auto visual_projection = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
 
-        auto x = vision_model->forward(ctx, pixel_values);  // [N, hidden_size]
-        x      = visual_projection->forward(ctx, x);        // [N, projection_dim]
+        auto x = vision_model->forward(ctx, backend, pixel_values, return_pooled, clip_skip);  // [N, hidden_size] or [N, n_token, hidden_size]
 
-        return x;  // [N, projection_dim]
+        if (return_pooled) {
+            x = visual_projection->forward(ctx, x);  // [N, projection_dim]
+        }
+
+        return x;
     }
 };
 
@@ -868,12 +884,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
     CLIPTextModel model;
 
     CLIPTextModelRunner(ggml_backend_t backend,
-                        std::map<std::string, enum ggml_type>& tensor_types,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types,
                         const std::string prefix,
                         CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                         bool with_final_ln  = true,
-                        int clip_skip_value = -1)
-        : GGMLRunner(backend), model(version, with_final_ln, clip_skip_value) {
+                        bool force_clip_f32 = false)
+        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, force_clip_f32) {
         model.init(params_ctx, tensor_types, prefix);
     }
 
@@ -881,19 +898,17 @@ struct CLIPTextModelRunner : public GGMLRunner {
         return "clip";
     }
 
-    void set_clip_skip(int clip_skip) {
-        model.set_clip_skip(clip_skip);
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* embeddings,
                                 size_t max_token_idx = 0,
-                                bool return_pooled   = false) {
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
         if (input_ids->ne[0] > model.n_token) {
@@ -901,14 +916,15 @@ struct CLIPTextModelRunner : public GGMLRunner {
             input_ids = ggml_reshape_2d(ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
         }
 
-        return model.forward(ctx, input_ids, embeddings, max_token_idx, return_pooled);
+        return model.forward(ctx, backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
     }
 
     struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
                                     int num_custom_embeddings    = 0,
                                     void* custom_embeddings_data = NULL,
                                     size_t max_token_idx         = 0,
-                                    bool return_pooled           = false) {
+                                    bool return_pooled           = false,
+                                    int clip_skip                = -1) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         input_ids = to_backend(input_ids);
@@ -927,7 +943,7 @@ struct CLIPTextModelRunner : public GGMLRunner {
             embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
         }
 
-        struct ggml_tensor* hidden_states = forward(compute_ctx, input_ids, embeddings, max_token_idx, return_pooled);
+        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -940,10 +956,11 @@ struct CLIPTextModelRunner : public GGMLRunner {
                  void* custom_embeddings_data,
                  size_t max_token_idx,
                  bool return_pooled,
+                 int clip_skip,
                  ggml_tensor** output,
                  ggml_context* output_ctx = NULL) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled);
+            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
         };
         GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
     }

common.hpp

@@ -182,9 +182,9 @@ protected:
     int64_t dim_in;
     int64_t dim_out;
 
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
-        enum ggml_type wtype      = (tensor_types.find(prefix + "proj.weight") != tensor_types.end()) ? tensor_types[prefix + "proj.weight"] : GGML_TYPE_F32;
-        enum ggml_type bias_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "proj.bias") != tensor_types.end()) ? tensor_types[prefix + "proj.bias"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+        enum ggml_type wtype      = get_type(prefix + "proj.weight", tensor_types, GGML_TYPE_F32);
+        enum ggml_type bias_wtype = GGML_TYPE_F32;
         params["proj.weight"]     = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
         params["proj.bias"]       = ggml_new_tensor_1d(ctx, bias_wtype, dim_out * 2);
     }
@@ -270,7 +270,10 @@ public:
         // to_out_1 is nn.Dropout(), skip for inference
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* context) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
         // x: [N, n_token, query_dim]
         // context: [N, n_context, context_dim]
         // return: [N, n_token, query_dim]
@@ -288,7 +291,7 @@ public:
         auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
         auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
 
-        x = ggml_nn_attention_ext(ctx, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]
+        x = ggml_nn_attention_ext(ctx, backend, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]
 
         x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
         return x;
@@ -327,7 +330,10 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* context) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
         // x: [N, n_token, query_dim]
         // context: [N, n_context, context_dim]
         // return: [N, n_token, query_dim]
@@ -352,11 +358,11 @@ public:
 
         auto r = x;
         x      = norm1->forward(ctx, x);
-        x      = attn1->forward(ctx, x, x);  // self-attention
+        x      = attn1->forward(ctx, backend, x, x);  // self-attention
         x      = ggml_add(ctx, x, r);
         r      = x;
         x      = norm2->forward(ctx, x);
-        x      = attn2->forward(ctx, x, context);  // cross-attention
+        x      = attn2->forward(ctx, backend, x, context);  // cross-attention
         x      = ggml_add(ctx, x, r);
         r      = x;
         x      = norm3->forward(ctx, x);
@@ -401,7 +407,10 @@ public:
         blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
     }
 
-    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* context) {
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx,
+                                        ggml_backend_t backend,
+                                        struct ggml_tensor* x,
+                                        struct ggml_tensor* context) {
         // x: [N, in_channels, h, w]
         // context: [N, max_position(aka n_token), hidden_size(aka context_dim)]
         auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
@@ -424,7 +433,7 @@ public:
             std::string name       = "transformer_blocks." + std::to_string(i);
             auto transformer_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[name]);
 
-            x = transformer_block->forward(ctx, x, context);
+            x = transformer_block->forward(ctx, backend, x, context);
         }
 
         x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
@@ -440,9 +449,9 @@ public:
 
 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
         // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
-        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.ypes.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
+        enum ggml_type wtype = GGML_TYPE_F32;
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 

conditioner.hpp

@@ -21,12 +21,12 @@ struct Conditioner {
                                               int clip_skip,
                                               int width,
                                               int height,
-                                              int adm_in_channels        = -1,
-                                              bool force_zero_embeddings = false)                                             = 0;
-    virtual void alloc_params_buffer()                                                                                        = 0;
-    virtual void free_params_buffer()                                                                                         = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors)                                       = 0;
-    virtual size_t get_params_buffer_size()                                                                                   = 0;
+                                              int adm_in_channels  = -1,
+                                              bool zero_out_masked = false)                                             = 0;
+    virtual void alloc_params_buffer()                                                                                  = 0;
+    virtual void free_params_buffer()                                                                                   = 0;
+    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors)                                 = 0;
+    virtual size_t get_params_buffer_size()                                                                             = 0;
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
                                                                                           int n_threads,
                                                                                           const std::string& text,
@@ -34,10 +34,10 @@ struct Conditioner {
                                                                                           int width,
                                                                                           int height,
                                                                                           int num_input_imgs,
-                                                                                          int adm_in_channels        = -1,
-                                                                                          bool force_zero_embeddings = false) = 0;
+                                                                                          int adm_in_channels  = -1,
+                                                                                          bool zero_out_masked = false) = 0;
     virtual std::string remove_trigger_from_prompt(ggml_context* work_ctx,
-                                                   const std::string& prompt)                                                 = 0;
+                                                   const std::string& prompt)                                           = 0;
 };
 
 // ldm.modules.encoders.modules.FrozenCLIPEmbedder
@@ -57,33 +57,20 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     std::vector<std::string> readed_embeddings;
 
     FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
-                                      std::map<std::string, enum ggml_type>& tensor_types,
+                                      bool offload_params_to_cpu,
+                                      const String2GGMLType& tensor_types,
                                       const std::string& embd_dir,
                                       SDVersion version = VERSION_SD1,
-                                      PMVersion pv      = PM_VERSION_1,
-                                      int clip_skip     = -1)
+                                      PMVersion pv      = PM_VERSION_1)
         : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
+        bool force_clip_f32 = embd_dir.size() > 0;
         if (sd_version_is_sd1(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
         } else if (sd_version_is_sd2(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
         } else if (sd_version_is_sdxl(version)) {
-            text_model  = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
-        }
-        set_clip_skip(clip_skip);
-    }
-
-    void set_clip_skip(int clip_skip) {
-        if (clip_skip <= 0) {
-            clip_skip = 1;
-            if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) {
-                clip_skip = 2;
-            }
-        }
-        text_model->set_clip_skip(clip_skip);
-        if (sd_version_is_sdxl(version)) {
-            text_model2->set_clip_skip(clip_skip);
+            text_model  = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
+            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
         }
     }
 
@@ -128,7 +115,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             return true;
         }
         struct ggml_init_params params;
-        params.mem_size               = 10 * 1024 * 1024;  // max for custom embeddings 10 MB
+        params.mem_size               = 100 * 1024 * 1024;  // max for custom embeddings 100 MB
         params.mem_buffer             = NULL;
         params.no_alloc               = false;
         struct ggml_context* embd_ctx = ggml_init(params);
@@ -154,7 +141,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             }
             return true;
         };
-        model_loader.load_tensors(on_load, NULL);
+        model_loader.load_tensors(on_load);
         readed_embeddings.push_back(embd_name);
         if (embd) {
             int64_t hidden_size = text_model->model.hidden_size;
@@ -409,9 +396,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                              int clip_skip,
                                              int width,
                                              int height,
-                                             int adm_in_channels        = -1,
-                                             bool force_zero_embeddings = false) {
-        set_clip_skip(clip_skip);
+                                             int adm_in_channels  = -1,
+                                             bool zero_out_masked = false) {
         int64_t t0                               = ggml_time_ms();
         struct ggml_tensor* hidden_states        = NULL;  // [N, n_token, hidden_size]
         struct ggml_tensor* chunk_hidden_states  = NULL;  // [n_token, hidden_size] or [n_token, hidden_size + hidden_size2]
@@ -420,6 +406,10 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         struct ggml_tensor* pooled               = NULL;
         std::vector<float> hidden_states_vec;
 
+        if (clip_skip <= 0) {
+            clip_skip = (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) ? 2 : 1;
+        }
+
         size_t chunk_len   = 77;
         size_t chunk_count = tokens.size() / chunk_len;
         for (int chunk_idx = 0; chunk_idx < chunk_count; chunk_idx++) {
@@ -454,6 +444,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                     token_embed_custom.data(),
                                     max_token_idx,
                                     false,
+                                    clip_skip,
                                     &chunk_hidden_states1,
                                     work_ctx);
                 if (sd_version_is_sdxl(version)) {
@@ -463,6 +454,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                          token_embed_custom.data(),
                                          max_token_idx,
                                          false,
+                                         clip_skip,
                                          &chunk_hidden_states2, work_ctx);
                     // concat
                     chunk_hidden_states = ggml_tensor_concat(work_ctx, chunk_hidden_states1, chunk_hidden_states2, 0);
@@ -474,6 +466,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                              token_embed_custom.data(),
                                              max_token_idx,
                                              true,
+                                             clip_skip,
                                              &pooled,
                                              work_ctx);
                     }
@@ -499,7 +492,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                 float new_mean = ggml_tensor_mean(result);
                 ggml_tensor_scale(result, (original_mean / new_mean));
             }
-            if (force_zero_embeddings) {
+            if (zero_out_masked) {
                 float* vec = (float*)result->data;
                 for (int i = 0; i < ggml_nelements(result); i++) {
                     vec[i] = 0;
@@ -562,8 +555,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                        int width,
                                        int height,
                                        int num_input_imgs,
-                                       int adm_in_channels        = -1,
-                                       bool force_zero_embeddings = false) {
+                                       int adm_in_channels  = -1,
+                                       bool zero_out_masked = false) {
         auto image_tokens = convert_token_to_id(trigger_word);
         // if(image_tokens.size() == 1){
         //     printf(" image token id is: %d \n", image_tokens[0]);
@@ -584,7 +577,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         // for(int i = 0; i < clsm.size(); ++i)
         //    printf("%d ", clsm[i]?1:0);
         // printf("\n");
-        auto cond = get_learned_condition_common(work_ctx, n_threads, tokens, weights, clip_skip, width, height, adm_in_channels, force_zero_embeddings);
+        auto cond = get_learned_condition_common(work_ctx, n_threads, tokens, weights, clip_skip, width, height, adm_in_channels, zero_out_masked);
         return std::make_tuple(cond, clsm);
     }
 
@@ -606,20 +599,22 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                       int clip_skip,
                                       int width,
                                       int height,
-                                      int adm_in_channels        = -1,
-                                      bool force_zero_embeddings = false) {
+                                      int adm_in_channels  = -1,
+                                      bool zero_out_masked = false) {
         auto tokens_and_weights     = tokenize(text, true);
         std::vector<int>& tokens    = tokens_and_weights.first;
         std::vector<float>& weights = tokens_and_weights.second;
-        return get_learned_condition_common(work_ctx, n_threads, tokens, weights, clip_skip, width, height, adm_in_channels, force_zero_embeddings);
+        return get_learned_condition_common(work_ctx, n_threads, tokens, weights, clip_skip, width, height, adm_in_channels, zero_out_masked);
     }
 };
 
 struct FrozenCLIPVisionEmbedder : public GGMLRunner {
     CLIPVisionModelProjection vision_model;
 
-    FrozenCLIPVisionEmbedder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
-        : vision_model(OPEN_CLIP_VIT_H_14, true), GGMLRunner(backend) {
+    FrozenCLIPVisionEmbedder(ggml_backend_t backend,
+                             bool offload_params_to_cpu,
+                             const String2GGMLType& tensor_types = {})
+        : vision_model(OPEN_CLIP_VIT_H_14), GGMLRunner(backend, offload_params_to_cpu) {
         vision_model.init(params_ctx, tensor_types, "cond_stage_model.transformer");
     }
 
@@ -631,12 +626,12 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
         vision_model.get_param_tensors(tensors, "cond_stage_model.transformer");
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* pixel_values) {
+    struct ggml_cgraph* build_graph(struct ggml_tensor* pixel_values, bool return_pooled, int clip_skip) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         pixel_values = to_backend(pixel_values);
 
-        struct ggml_tensor* hidden_states = vision_model.forward(compute_ctx, pixel_values);
+        struct ggml_tensor* hidden_states = vision_model.forward(compute_ctx, runtime_backend, pixel_values, return_pooled, clip_skip);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -645,10 +640,12 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
 
     void compute(const int n_threads,
                  ggml_tensor* pixel_values,
+                 bool return_pooled,
+                 int clip_skip,
                  ggml_tensor** output,
                  ggml_context* output_ctx) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(pixel_values);
+            return build_graph(pixel_values, return_pooled, clip_skip);
         };
         GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
     }
@@ -663,21 +660,12 @@ struct SD3CLIPEmbedder : public Conditioner {
     std::shared_ptr<T5Runner> t5;
 
     SD3CLIPEmbedder(ggml_backend_t backend,
-                    std::map<std::string, enum ggml_type>& tensor_types,
-                    int clip_skip = -1)
+                    bool offload_params_to_cpu,
+                    const String2GGMLType& tensor_types = {})
         : clip_g_tokenizer(0) {
-        clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
-        clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
-        t5     = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
-        set_clip_skip(clip_skip);
-    }
-
-    void set_clip_skip(int clip_skip) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l->set_clip_skip(clip_skip);
-        clip_g->set_clip_skip(clip_skip);
+        clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
+        clip_g = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
+        t5     = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.t5xxl.transformer");
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
@@ -773,8 +761,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                              int n_threads,
                                              std::vector<std::pair<std::vector<int>, std::vector<float>>> token_and_weights,
                                              int clip_skip,
-                                             bool force_zero_embeddings = false) {
-        set_clip_skip(clip_skip);
+                                             bool zero_out_masked = false) {
         auto& clip_l_tokens  = token_and_weights[0].first;
         auto& clip_l_weights = token_and_weights[0].second;
         auto& clip_g_tokens  = token_and_weights[1].first;
@@ -782,6 +769,10 @@ struct SD3CLIPEmbedder : public Conditioner {
         auto& t5_tokens      = token_and_weights[2].first;
         auto& t5_weights     = token_and_weights[2].second;
 
+        if (clip_skip <= 0) {
+            clip_skip = 2;
+        }
+
         int64_t t0                                 = ggml_time_ms();
         struct ggml_tensor* hidden_states          = NULL;  // [N, n_token*2, 4096]
         struct ggml_tensor* chunk_hidden_states    = NULL;  // [n_token*2, 4096]
@@ -812,6 +803,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                 NULL,
                                 max_token_idx,
                                 false,
+                                clip_skip,
                                 &chunk_hidden_states_l,
                                 work_ctx);
                 {
@@ -839,6 +831,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                     NULL,
                                     max_token_idx,
                                     true,
+                                    clip_skip,
                                     &pooled_l,
                                     work_ctx);
                 }
@@ -860,6 +853,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                 NULL,
                                 max_token_idx,
                                 false,
+                                clip_skip,
                                 &chunk_hidden_states_g,
                                 work_ctx);
 
@@ -888,6 +882,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                     NULL,
                                     max_token_idx,
                                     true,
+                                    clip_skip,
                                     &pooled_g,
                                     work_ctx);
                 }
@@ -952,7 +947,7 @@ struct SD3CLIPEmbedder : public Conditioner {
 
             int64_t t1 = ggml_time_ms();
             LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
-            if (force_zero_embeddings) {
+            if (zero_out_masked) {
                 float* vec = (float*)chunk_hidden_states->data;
                 for (int i = 0; i < ggml_nelements(chunk_hidden_states); i++) {
                     vec[i] = 0;
@@ -978,10 +973,10 @@ struct SD3CLIPEmbedder : public Conditioner {
                                       int clip_skip,
                                       int width,
                                       int height,
-                                      int adm_in_channels        = -1,
-                                      bool force_zero_embeddings = false) {
+                                      int adm_in_channels  = -1,
+                                      bool zero_out_masked = false) {
         auto tokens_and_weights = tokenize(text, 77, true);
-        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, force_zero_embeddings);
+        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
 
     std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
@@ -991,8 +986,8 @@ struct SD3CLIPEmbedder : public Conditioner {
                                                                                   int width,
                                                                                   int height,
                                                                                   int num_input_imgs,
-                                                                                  int adm_in_channels        = -1,
-                                                                                  bool force_zero_embeddings = false) {
+                                                                                  int adm_in_channels  = -1,
+                                                                                  bool zero_out_masked = false) {
         GGML_ASSERT(0 && "Not implemented yet!");
     }
 
@@ -1010,18 +1005,10 @@ struct FluxCLIPEmbedder : public Conditioner {
     size_t chunk_len = 256;
 
     FluxCLIPEmbedder(ggml_backend_t backend,
-                     std::map<std::string, enum ggml_type>& tensor_types,
-                     int clip_skip = -1) {
-        clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
-        t5     = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
-        set_clip_skip(clip_skip);
-    }
-
-    void set_clip_skip(int clip_skip) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l->set_clip_skip(clip_skip);
+                     bool offload_params_to_cpu,
+                     const String2GGMLType& tensor_types = {}) {
+        clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
+        t5     = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.t5xxl.transformer");
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
@@ -1101,13 +1088,16 @@ struct FluxCLIPEmbedder : public Conditioner {
                                              int n_threads,
                                              std::vector<std::pair<std::vector<int>, std::vector<float>>> token_and_weights,
                                              int clip_skip,
-                                             bool force_zero_embeddings = false) {
-        set_clip_skip(clip_skip);
+                                             bool zero_out_masked = false) {
         auto& clip_l_tokens  = token_and_weights[0].first;
         auto& clip_l_weights = token_and_weights[0].second;
         auto& t5_tokens      = token_and_weights[1].first;
         auto& t5_weights     = token_and_weights[1].second;
 
+        if (clip_skip <= 0) {
+            clip_skip = 2;
+        }
+
         int64_t t0                              = ggml_time_ms();
         struct ggml_tensor* hidden_states       = NULL;  // [N, n_token, 4096]
         struct ggml_tensor* chunk_hidden_states = NULL;  // [n_token, 4096]
@@ -1136,6 +1126,7 @@ struct FluxCLIPEmbedder : public Conditioner {
                                 NULL,
                                 max_token_idx,
                                 true,
+                                clip_skip,
                                 &pooled,
                                 work_ctx);
             }
@@ -1173,7 +1164,7 @@ struct FluxCLIPEmbedder : public Conditioner {
 
             int64_t t1 = ggml_time_ms();
             LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
-            if (force_zero_embeddings) {
+            if (zero_out_masked) {
                 float* vec = (float*)chunk_hidden_states->data;
                 for (int i = 0; i < ggml_nelements(chunk_hidden_states); i++) {
                     vec[i] = 0;
@@ -1199,10 +1190,10 @@ struct FluxCLIPEmbedder : public Conditioner {
                                       int clip_skip,
                                       int width,
                                       int height,
-                                      int adm_in_channels        = -1,
-                                      bool force_zero_embeddings = false) {
+                                      int adm_in_channels  = -1,
+                                      bool zero_out_masked = false) {
         auto tokens_and_weights = tokenize(text, chunk_len, true);
-        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, force_zero_embeddings);
+        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
 
     std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
@@ -1212,8 +1203,8 @@ struct FluxCLIPEmbedder : public Conditioner {
                                                                                   int width,
                                                                                   int height,
                                                                                   int num_input_imgs,
-                                                                                  int adm_in_channels        = -1,
-                                                                                  bool force_zero_embeddings = false) {
+                                                                                  int adm_in_channels  = -1,
+                                                                                  bool zero_out_masked = false) {
         GGML_ASSERT(0 && "Not implemented yet!");
     }
 
@@ -1223,23 +1214,22 @@ struct FluxCLIPEmbedder : public Conditioner {
     }
 };
 
-struct PixArtCLIPEmbedder : public Conditioner {
+struct T5CLIPEmbedder : public Conditioner {
     T5UniGramTokenizer t5_tokenizer;
     std::shared_ptr<T5Runner> t5;
     size_t chunk_len = 512;
     bool use_mask    = false;
     int mask_pad     = 1;
+    bool is_umt5     = false;
 
-    PixArtCLIPEmbedder(ggml_backend_t backend,
-                       std::map<std::string, enum ggml_type>& tensor_types,
-                       int clip_skip = -1,
-                       bool use_mask = false,
-                       int mask_pad  = 1)
-        : use_mask(use_mask), mask_pad(mask_pad) {
-        t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
-    }
-
-    void set_clip_skip(int clip_skip) {
+    T5CLIPEmbedder(ggml_backend_t backend,
+                   bool offload_params_to_cpu,
+                   const String2GGMLType& tensor_types = {},
+                   bool use_mask                       = false,
+                   int mask_pad                        = 1,
+                   bool is_umt5                        = false)
+        : use_mask(use_mask), mask_pad(mask_pad), t5_tokenizer(is_umt5) {
+        t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.t5xxl.transformer", is_umt5);
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
@@ -1317,16 +1307,16 @@ struct PixArtCLIPEmbedder : public Conditioner {
                                              int n_threads,
                                              std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> token_and_weights,
                                              int clip_skip,
-                                             bool force_zero_embeddings = false) {
+                                             bool zero_out_masked = false) {
         auto& t5_tokens        = std::get<0>(token_and_weights);
         auto& t5_weights       = std::get<1>(token_and_weights);
         auto& t5_attn_mask_vec = std::get<2>(token_and_weights);
 
         int64_t t0                              = ggml_time_ms();
-        struct ggml_tensor* hidden_states       = NULL;                                               // [N, n_token, 4096]
-        struct ggml_tensor* chunk_hidden_states = NULL;                                               // [n_token, 4096]
-        struct ggml_tensor* pooled              = NULL;                                               // [768,]
-        struct ggml_tensor* t5_attn_mask        = vector_to_ggml_tensor(work_ctx, t5_attn_mask_vec);  // [768,]
+        struct ggml_tensor* hidden_states       = NULL;  // [N, n_token, 4096]
+        struct ggml_tensor* chunk_hidden_states = NULL;  // [n_token, 4096]
+        struct ggml_tensor* pooled              = NULL;
+        struct ggml_tensor* t5_attn_mask        = vector_to_ggml_tensor(work_ctx, t5_attn_mask_vec);  // [n_token]
 
         std::vector<float> hidden_states_vec;
 
@@ -1367,10 +1357,16 @@ struct PixArtCLIPEmbedder : public Conditioner {
 
             int64_t t1 = ggml_time_ms();
             LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
-            if (force_zero_embeddings) {
-                float* vec = (float*)chunk_hidden_states->data;
-                for (int i = 0; i < ggml_nelements(chunk_hidden_states); i++) {
-                    vec[i] = 0;
+            if (zero_out_masked) {
+                auto tensor = chunk_hidden_states;
+                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                            if (chunk_mask[i1] < 0.f) {
+                                ggml_tensor_set_f32(tensor, 0.f, i0, i1, i2);
+                            }
+                        }
+                    }
                 }
             }
 
@@ -1379,16 +1375,12 @@ struct PixArtCLIPEmbedder : public Conditioner {
                                      ((float*)chunk_hidden_states->data) + ggml_nelements(chunk_hidden_states));
         }
 
-        if (hidden_states_vec.size() > 0) {
-            hidden_states = vector_to_ggml_tensor(work_ctx, hidden_states_vec);
-            hidden_states = ggml_reshape_2d(work_ctx,
-                                            hidden_states,
-                                            chunk_hidden_states->ne[0],
-                                            ggml_nelements(hidden_states) / chunk_hidden_states->ne[0]);
-        } else {
-            hidden_states = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 4096, 256);
-            ggml_set_f32(hidden_states, 0.f);
-        }
+        GGML_ASSERT(hidden_states_vec.size() > 0);
+        hidden_states = vector_to_ggml_tensor(work_ctx, hidden_states_vec);
+        hidden_states = ggml_reshape_2d(work_ctx,
+                                        hidden_states,
+                                        chunk_hidden_states->ne[0],
+                                        ggml_nelements(hidden_states) / chunk_hidden_states->ne[0]);
 
         modify_mask_to_attend_padding(t5_attn_mask, ggml_nelements(t5_attn_mask), mask_pad);
 
@@ -1401,10 +1393,10 @@ struct PixArtCLIPEmbedder : public Conditioner {
                                       int clip_skip,
                                       int width,
                                       int height,
-                                      int adm_in_channels        = -1,
-                                      bool force_zero_embeddings = false) {
+                                      int adm_in_channels  = -1,
+                                      bool zero_out_masked = false) {
         auto tokens_and_weights = tokenize(text, chunk_len, true);
-        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, force_zero_embeddings);
+        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
 
     std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
@@ -1414,8 +1406,8 @@ struct PixArtCLIPEmbedder : public Conditioner {
                                                                                   int width,
                                                                                   int height,
                                                                                   int num_input_imgs,
-                                                                                  int adm_in_channels        = -1,
-                                                                                  bool force_zero_embeddings = false) {
+                                                                                  int adm_in_channels  = -1,
+                                                                                  bool zero_out_masked = false) {
         GGML_ASSERT(0 && "Not implemented yet!");
     }
 

control.hpp

@@ -174,10 +174,11 @@ public:
 
     struct ggml_tensor* attention_layer_forward(std::string name,
                                                 struct ggml_context* ctx,
+                                                ggml_backend_t backend,
                                                 struct ggml_tensor* x,
                                                 struct ggml_tensor* context) {
         auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
-        return block->forward(ctx, x, context);
+        return block->forward(ctx, backend, x, context);
     }
 
     struct ggml_tensor* input_hint_block_forward(struct ggml_context* ctx,
@@ -199,6 +200,7 @@ public:
     }
 
     std::vector<struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                             ggml_backend_t backend,
                                              struct ggml_tensor* x,
                                              struct ggml_tensor* hint,
                                              struct ggml_tensor* guided_hint,
@@ -272,7 +274,7 @@ public:
                 h                = resblock_forward(name, ctx, h, emb);  // [N, mult*model_channels, h, w]
                 if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
                     std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
-                    h                = attention_layer_forward(name, ctx, h, context);  // [N, mult*model_channels, h, w]
+                    h                = attention_layer_forward(name, ctx, backend, h, context);  // [N, mult*model_channels, h, w]
                 }
 
                 auto zero_conv = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs." + std::to_string(input_block_idx) + ".0"]);
@@ -296,9 +298,9 @@ public:
         // [N, 4*model_channels, h/8, w/8]
 
         // middle_block
-        h = resblock_forward("middle_block.0", ctx, h, emb);             // [N, 4*model_channels, h/8, w/8]
-        h = attention_layer_forward("middle_block.1", ctx, h, context);  // [N, 4*model_channels, h/8, w/8]
-        h = resblock_forward("middle_block.2", ctx, h, emb);             // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.0", ctx, h, emb);                      // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, backend, h, context);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb);                      // [N, 4*model_channels, h/8, w/8]
 
         // out
         outs.push_back(middle_block_out->forward(ctx, h));
@@ -317,12 +319,24 @@ struct ControlNet : public GGMLRunner {
     bool guided_hint_cached         = false;
 
     ControlNet(ggml_backend_t backend,
-               std::map<std::string, enum ggml_type>& tensor_types,
-               SDVersion version = VERSION_SD1)
-        : GGMLRunner(backend), control_net(version) {
+               bool offload_params_to_cpu,
+               const String2GGMLType& tensor_types = {},
+               SDVersion version                   = VERSION_SD1)
+        : GGMLRunner(backend, offload_params_to_cpu), control_net(version) {
         control_net.init(params_ctx, tensor_types, "");
     }
 
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        control_net.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
     ~ControlNet() {
         free_control_ctx();
     }
@@ -346,7 +360,7 @@ struct ControlNet : public GGMLRunner {
             control_buffer_size += ggml_nbytes(controls[i]);
         }
 
-        control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, backend);
+        control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, runtime_backend);
 
         LOG_DEBUG("control buffer size %.2fMB", control_buffer_size * 1.f / 1024.f / 1024.f);
     }
@@ -391,6 +405,7 @@ struct ControlNet : public GGMLRunner {
         timesteps = to_backend(timesteps);
 
         auto outs = control_net.forward(compute_ctx,
+                                        runtime_backend,
                                         x,
                                         hint,
                                         guided_hint_cached ? guided_hint : NULL,
@@ -443,7 +458,7 @@ struct ControlNet : public GGMLRunner {
             return false;
         }
 
-        bool success = model_loader.load_tensors(tensors, backend, ignore_tensors);
+        bool success = model_loader.load_tensors(tensors, ignore_tensors);
 
         if (!success) {
             LOG_ERROR("load control net tensors from model loader failed");

denoiser.hpp

@@ -251,8 +251,37 @@ struct KarrasSchedule : SigmaSchedule {
     }
 };
 
+// Close to Beta Schedule, but increadably simple in code.
+struct SmoothStepSchedule : SigmaSchedule {
+    static constexpr float smoothstep(float x) {
+        return x * x * (3.0f - 2.0f * 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((float)t_max));
+            result.push_back(0.f);
+            return result;
+        }
+
+        for (uint32_t i = 0; i < n; i++) {
+            float u = 1.f - float(i) / float(n);
+            result.push_back(t_to_sigma(std::round(smoothstep(u) * t_max)));
+        }
+
+        result.push_back(0.f);
+        return result;
+    }
+};
+
 struct Denoiser {
-    std::shared_ptr<SigmaSchedule> schedule                                                  = std::make_shared<DiscreteSchedule>();
+    std::shared_ptr<SigmaSchedule> scheduler                                                 = std::make_shared<DiscreteSchedule>();
     virtual float sigma_min()                                                                = 0;
     virtual float sigma_max()                                                                = 0;
     virtual float sigma_to_t(float sigma)                                                    = 0;
@@ -263,7 +292,7 @@ struct Denoiser {
 
     virtual std::vector<float> get_sigmas(uint32_t n) {
         auto bound_t_to_sigma = std::bind(&Denoiser::t_to_sigma, this, std::placeholders::_1);
-        return schedule->get_sigmas(n, sigma_min(), sigma_max(), bound_t_to_sigma);
+        return scheduler->get_sigmas(n, sigma_min(), sigma_max(), bound_t_to_sigma);
     }
 };
 
@@ -349,7 +378,7 @@ struct EDMVDenoiser : public CompVisVDenoiser {
 
     EDMVDenoiser(float min_sigma = 0.002, float max_sigma = 120.0)
         : min_sigma(min_sigma), max_sigma(max_sigma) {
-        schedule = std::make_shared<ExponentialSchedule>();
+        scheduler = std::make_shared<ExponentialSchedule>();
     }
 
     float t_to_sigma(float t) {
@@ -382,7 +411,8 @@ struct DiscreteFlowDenoiser : public Denoiser {
 
     float sigma_data = 1.0f;
 
-    DiscreteFlowDenoiser() {
+    DiscreteFlowDenoiser(float shift = 3.0f)
+        : shift(shift) {
         set_parameters();
     }
 

diffusion_model.hpp

@@ -4,22 +4,31 @@
 #include "flux.hpp"
 #include "mmdit.hpp"
 #include "unet.hpp"
+#include "wan.hpp"
+
+struct DiffusionParams {
+    struct ggml_tensor* x                     = NULL;
+    struct ggml_tensor* timesteps             = NULL;
+    struct ggml_tensor* context               = NULL;
+    struct ggml_tensor* c_concat              = NULL;
+    struct ggml_tensor* y                     = NULL;
+    struct ggml_tensor* guidance              = NULL;
+    std::vector<ggml_tensor*> ref_latents     = {};
+    bool increase_ref_index                   = false;
+    int num_video_frames                      = -1;
+    std::vector<struct ggml_tensor*> controls = {};
+    float control_strength                    = 0.f;
+    struct ggml_tensor* vace_context          = NULL;
+    float vace_strength                       = 1.f;
+    std::vector<int> skip_layers              = {};
+};
 
 struct DiffusionModel {
+    virtual std::string get_desc()                                                      = 0;
     virtual void compute(int n_threads,
-                         struct ggml_tensor* x,
-                         struct ggml_tensor* timesteps,
-                         struct ggml_tensor* context,
-                         struct ggml_tensor* c_concat,
-                         struct ggml_tensor* y,
-                         struct ggml_tensor* guidance,
-                         std::vector<ggml_tensor*> ref_latents     = {},
-                         int num_video_frames                      = -1,
-                         std::vector<struct ggml_tensor*> controls = {},
-                         float control_strength                    = 0.f,
-                         struct ggml_tensor** output               = NULL,
-                         struct ggml_context* output_ctx           = NULL,
-                         std::vector<int> skip_layers              = std::vector<int>())             = 0;
+                         DiffusionParams diffusion_params,
+                         struct ggml_tensor** output     = NULL,
+                         struct ggml_context* output_ctx = NULL)                        = 0;
     virtual void alloc_params_buffer()                                                  = 0;
     virtual void free_params_buffer()                                                   = 0;
     virtual void free_compute_buffer()                                                  = 0;
@@ -32,10 +41,15 @@ struct UNetModel : public DiffusionModel {
     UNetModelRunner unet;
 
     UNetModel(ggml_backend_t backend,
-              std::map<std::string, enum ggml_type>& tensor_types,
-              SDVersion version = VERSION_SD1,
-              bool flash_attn   = false)
-        : unet(backend, tensor_types, "model.diffusion_model", version, flash_attn) {
+              bool offload_params_to_cpu,
+              const String2GGMLType& tensor_types = {},
+              SDVersion version                   = VERSION_SD1,
+              bool flash_attn                     = false)
+        : unet(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return unet.get_desc();
     }
 
     void alloc_params_buffer() {
@@ -63,21 +77,18 @@ struct UNetModel : public DiffusionModel {
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 int num_video_frames                      = -1,
-                 std::vector<struct ggml_tensor*> controls = {},
-                 float control_strength                    = 0.f,
-                 struct ggml_tensor** output               = NULL,
-                 struct ggml_context* output_ctx           = NULL,
-                 std::vector<int> skip_layers              = std::vector<int>()) {
-        (void)skip_layers;  // SLG doesn't work with UNet models
-        return unet.compute(n_threads, x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength, output, output_ctx);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return unet.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.num_video_frames,
+                            diffusion_params.controls,
+                            diffusion_params.control_strength, output, output_ctx);
     }
 };
 
@@ -85,8 +96,14 @@ struct MMDiTModel : public DiffusionModel {
     MMDiTRunner mmdit;
 
     MMDiTModel(ggml_backend_t backend,
-               std::map<std::string, enum ggml_type>& tensor_types)
-        : mmdit(backend, tensor_types, "model.diffusion_model") {
+               bool offload_params_to_cpu,
+               bool flash_attn                     = false,
+               const String2GGMLType& tensor_types = {})
+        : mmdit(backend, offload_params_to_cpu, flash_attn, tensor_types, "model.diffusion_model") {
+    }
+
+    std::string get_desc() {
+        return mmdit.get_desc();
     }
 
     void alloc_params_buffer() {
@@ -114,20 +131,17 @@ struct MMDiTModel : public DiffusionModel {
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 int num_video_frames                      = -1,
-                 std::vector<struct ggml_tensor*> controls = {},
-                 float control_strength                    = 0.f,
-                 struct ggml_tensor** output               = NULL,
-                 struct ggml_context* output_ctx           = NULL,
-                 std::vector<int> skip_layers              = std::vector<int>()) {
-        return mmdit.compute(n_threads, x, timesteps, context, y, output, output_ctx, skip_layers);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return mmdit.compute(n_threads,
+                             diffusion_params.x,
+                             diffusion_params.timesteps,
+                             diffusion_params.context,
+                             diffusion_params.y,
+                             output,
+                             output_ctx,
+                             diffusion_params.skip_layers);
     }
 };
 
@@ -135,11 +149,16 @@ struct FluxModel : public DiffusionModel {
     Flux::FluxRunner flux;
 
     FluxModel(ggml_backend_t backend,
-              std::map<std::string, enum ggml_type>& tensor_types,
-              SDVersion version = VERSION_FLUX,
-              bool flash_attn   = false,
-              bool use_mask     = false)
-        : flux(backend, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
+              bool offload_params_to_cpu,
+              const String2GGMLType& tensor_types = {},
+              SDVersion version                   = VERSION_FLUX,
+              bool flash_attn                     = false,
+              bool use_mask                       = false)
+        : flux(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
+    }
+
+    std::string get_desc() {
+        return flux.get_desc();
     }
 
     void alloc_params_buffer() {
@@ -167,20 +186,80 @@ struct FluxModel : public DiffusionModel {
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 int num_video_frames                      = -1,
-                 std::vector<struct ggml_tensor*> controls = {},
-                 float control_strength                    = 0.f,
-                 struct ggml_tensor** output               = NULL,
-                 struct ggml_context* output_ctx           = NULL,
-                 std::vector<int> skip_layers              = std::vector<int>()) {
-        return flux.compute(n_threads, x, timesteps, context, c_concat, y, guidance, ref_latents, output, output_ctx, skip_layers);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return flux.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.guidance,
+                            diffusion_params.ref_latents,
+                            diffusion_params.increase_ref_index,
+                            output,
+                            output_ctx,
+                            diffusion_params.skip_layers);
+    }
+};
+
+struct WanModel : public DiffusionModel {
+    std::string prefix;
+    WAN::WanRunner wan;
+
+    WanModel(ggml_backend_t backend,
+             bool offload_params_to_cpu,
+             const String2GGMLType& tensor_types = {},
+             const std::string prefix            = "model.diffusion_model",
+             SDVersion version                   = VERSION_WAN2,
+             bool flash_attn                     = false)
+        : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return wan.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        wan.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        wan.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        wan.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        wan.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() {
+        return wan.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return wan.compute(n_threads,
+                           diffusion_params.x,
+                           diffusion_params.timesteps,
+                           diffusion_params.context,
+                           diffusion_params.y,
+                           diffusion_params.c_concat,
+                           NULL,
+                           diffusion_params.vace_context,
+                           diffusion_params.vace_strength,
+                           output,
+                           output_ctx);
     }
 };
 

docs/chroma.md

@@ -24,7 +24,7 @@ You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](h
 For example:
 
 ```
- .\bin\Release\sd.exe -diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask
+ .\bin\Release\sd.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
 ```
 
 ![](../assets/flux/chroma_v40.png)

docs/flux.md

@@ -28,7 +28,7 @@ Using fp16 will lead to overflow, but ggml's support for bf16 is not yet fully d
 For example:
 
 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```
 
 Using formats of different precisions will yield results of varying quality.
@@ -44,7 +44,7 @@ Using formats of different precisions will yield results of varying quality.
 
 
 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
 ```
 
 | q8_0  |
@@ -60,7 +60,7 @@ Since many flux LoRA training libraries have used various LoRA naming formats, i
 - LoRA model from https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main (using comfy converted version!!!)
 
 ```
-.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models
+.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
 ```
 
 ![output](../assets/flux/flux1-dev-q8_0%20with%20lora.png)

docs/kontext.md

@@ -27,7 +27,7 @@ You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](ht
 For example:
 
 ```
- .\bin\Release\sd.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v
+ .\bin\Release\sd.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```
 
 

docs/lora.md

@@ -10,4 +10,30 @@ Here's a simple example:
 ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
 ```
 
-`../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
+`../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
+
+# Support matrix
+
+> ℹ️ CUDA `get_rows` support is defined here:  
+> [ggml-org/ggml/src/ggml-cuda/getrows.cu#L156](https://github.com/ggml-org/ggml/blob/7dee1d6a1e7611f238d09be96738388da97c88ed/src/ggml-cuda/getrows.cu#L156)  
+> Currently only the basic types + Q4/Q5/Q8 are implemented. K-quants are **not** supported.
+
+NOTE: The other backends may have different support.
+
+| Quant / Type | CUDA |
+|--------------|------|
+| F32          | ✔️   |
+| F16          | ✔️   |
+| BF16         | ✔️   |
+| I32          | ✔️   |
+| Q4_0         | ✔️   |
+| Q4_1         | ✔️   |
+| Q5_0         | ✔️   |
+| Q5_1         | ✔️   |
+| Q8_0         | ✔️   |
+| Q2_K         | ❌   |
+| Q3_K         | ❌   |
+| Q4_K         | ❌   |
+| Q5_K         | ❌   |
+| Q6_K         | ❌   |
+| Q8_K         | ❌   |

docs/photo_maker.md

@@ -6,16 +6,15 @@ You can use [PhotoMaker](https://github.com/TencentARC/PhotoMaker) to personaliz
 
 Download PhotoMaker model file (in safetensor format) [here](https://huggingface.co/bssrdf/PhotoMaker). The official release of the model file (in .bin format) does not work with ```stablediffusion.cpp```.
 
-- Specify the PhotoMaker model path using the `--stacked-id-embd-dir PATH` parameter.
-- Specify the input images path using the `--input-id-images-dir PATH` parameter.
-  - input images **must** have the same width and height for preprocessing (to be improved)
+- Specify the PhotoMaker model path using the `--photo-maker PATH` parameter.
+- Specify the input images path using the `--pm-id-images-dir PATH` parameter.
 
 In prompt, make sure you have a class word followed by the trigger word ```"img"``` (hard-coded for now). The class word could be one of ```"man, woman, girl, boy"```. If input ID images contain asian faces, add ```Asian``` before the class
 word.
 
 Another PhotoMaker specific parameter:
 
-- ```--style-ratio  (0-100)%```: default is 20 and 10-20 typically gets good results. Lower ratio means more faithfully following input ID (not necessarily better quality).
+- ```--pm-style-strength  (0-100)%```: default is 20 and 10-20 typically gets good results. Lower ratio means more faithfully following input ID (not necessarily better quality).
 
 Other parameters recommended for running Photomaker:
 
@@ -28,7 +27,7 @@ If on low memory GPUs (<= 8GB), recommend running with ```--vae-on-cpu``` option
 Example:
 
 ```bash
-bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --stacked-id-embd-dir ../models/photomaker-v1.safetensors --input-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --style-ratio 10 --vae-on-cpu -o output.png
+bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
 ```
 
 ## PhotoMaker Version 2

docs/sd3.md

@@ -14,7 +14,7 @@
 For example:
 
 ```
-.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v
+.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```
 
 ![](../assets/sd3.5_large.png)

docs/wan.md

@@ -0,0 +1,204 @@
+# How to Use
+
+## Download weights
+
+- Download Wan
+    - Wan2.1
+        - Wan2.1 T2V 1.3B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+        - Wan2.1 T2V 14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-T2V-14B-gguf/tree/main
+        - Wan2.1 I2V 14B 480P
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/tree/main
+        - Wan2.1 I2V 14B 720P
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf/tree/main
+        - Wan2.1 FLF2V 14B 720P
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-FLF2V-14B-720P-gguf/tree/main
+        - Wan2.1 VACE 1.3B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/calcuis/wan-1.3b-gguf/tree/main
+        - Wan2.1 VACE 14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.1_14B_VACE-GGUF/tree/main
+    - Wan2.2
+        - Wan2.2 TI2V 5B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-TI2V-5B-GGUF/tree/main
+        - Wan2.2 T2V A14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main
+        - Wan2.2 I2V A14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main
+- Download vae
+    - wan_2.1_vae (for all the wan model except Wan2.2 TI2V 5B)
+        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
+    - wan_2.2_vae (for Wan2.2 TI2V 5B only)
+        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan2.2_vae.safetensors
+- Download umt5_xxl
+    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp16.safetensors
+    - gguf: https://huggingface.co/city96/umt5-xxl-encoder-gguf/tree/main
+
+- Download clip_vison_h (for Wan2.1 I2V/FLF2V only)
+    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/clip_vision/clip_vision_h.safetensors
+
+
+## Examples
+
+### Wan2.1 T2V 1.3B
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 T2V 14B
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+### Wan2.1 I2V 14B
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 T2V A14B
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.2_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 I2V A14B
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.2_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 T2V A14B T2I
+
+```
+.\bin\Release\sd.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
+```
+
+<img width="832" height="480" alt="Wan2 2_14B_t2i" src="../assets/wan/Wan2.2_14B_t2i.png" />
+
+### Wan2.2 T2V 14B with Lora
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.2_14B_t2v_lora.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+### Wan2.2 TI2V 5B
+
+#### T2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.2_5B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+#### I2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.2_5B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 FLF2V 14B
+
+```
+.\bin\Release\sd.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
+```
+
+
+<video src=../assets/wan/Wan2.1_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 FLF2V 14B
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 VACE 1.3B
+
+#### T2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### R2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### V2V
+
+```
+mkdir post+depth
+ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 VACE 14B
+
+#### T2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### R2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+#### V2V
+
+```
+.\bin\Release\sd.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
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>

esrgan.hpp

@@ -142,11 +142,24 @@ struct ESRGAN : public GGMLRunner {
     int scale     = 4;
     int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
 
-    ESRGAN(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
-        : GGMLRunner(backend) {
+    ESRGAN(ggml_backend_t backend,
+           bool offload_params_to_cpu,
+           const String2GGMLType& tensor_types = {})
+        : GGMLRunner(backend, offload_params_to_cpu) {
         rrdb_net.init(params_ctx, tensor_types, "");
     }
 
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        rrdb_net.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
     std::string get_desc() {
         return "esrgan";
     }
@@ -164,7 +177,7 @@ struct ESRGAN : public GGMLRunner {
             return false;
         }
 
-        bool success = model_loader.load_tensors(esrgan_tensors, backend);
+        bool success = model_loader.load_tensors(esrgan_tensors);
 
         if (!success) {
             LOG_ERROR("load esrgan tensors from model loader failed");

examples/cli/CMakeLists.txt

@@ -3,4 +3,4 @@ set(TARGET sd)
 add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PUBLIC cxx_std_11)
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

examples/cli/avi_writer.h

@@ -0,0 +1,217 @@
+#ifndef __AVI_WRITER_H__
+#define __AVI_WRITER_H__
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "stable-diffusion.h"
+
+#ifndef INCLUDE_STB_IMAGE_WRITE_H
+#include "stb_image_write.h"
+#endif
+
+typedef struct {
+    uint32_t offset;
+    uint32_t size;
+} avi_index_entry;
+
+// Write 32-bit little-endian integer
+void write_u32_le(FILE* f, uint32_t val) {
+    fwrite(&val, 4, 1, f);
+}
+
+// Write 16-bit little-endian integer
+void write_u16_le(FILE* f, uint16_t val) {
+    fwrite(&val, 2, 1, f);
+}
+
+/**
+ * Create an MJPG AVI file from an array of sd_image_t images.
+ * Images are encoded to JPEG using stb_image_write.
+ *
+ * @param filename Output AVI file name.
+ * @param images Array of input images.
+ * @param num_images Number of images in the array.
+ * @param fps Frames per second for the video.
+ * @param quality JPEG quality (0-100).
+ * @return 0 on success, -1 on failure.
+ */
+int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality = 90) {
+    if (num_images == 0) {
+        fprintf(stderr, "Error: Image array is empty.\n");
+        return -1;
+    }
+
+    FILE* f = fopen(filename, "wb");
+    if (!f) {
+        perror("Error opening file for writing");
+        return -1;
+    }
+
+    uint32_t width    = images[0].width;
+    uint32_t height   = images[0].height;
+    uint32_t channels = images[0].channel;
+    if (channels != 3 && channels != 4) {
+        fprintf(stderr, "Error: Unsupported channel count: %u\n", channels);
+        fclose(f);
+        return -1;
+    }
+
+    // --- RIFF AVI Header ---
+    fwrite("RIFF", 4, 1, f);
+    long riff_size_pos = ftell(f);
+    write_u32_le(f, 0);  // Placeholder for file size
+    fwrite("AVI ", 4, 1, f);
+
+    // 'hdrl' LIST (header list)
+    fwrite("LIST", 4, 1, f);
+    write_u32_le(f, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
+    fwrite("hdrl", 4, 1, f);
+
+    // 'avih' chunk (AVI main header)
+    fwrite("avih", 4, 1, f);
+    write_u32_le(f, 56);
+    write_u32_le(f, 1000000 / fps);       // Microseconds per frame
+    write_u32_le(f, 0);                   // Max bytes per second
+    write_u32_le(f, 0);                   // Padding granularity
+    write_u32_le(f, 0x110);               // Flags (HASINDEX | ISINTERLEAVED)
+    write_u32_le(f, num_images);          // Total frames
+    write_u32_le(f, 0);                   // Initial frames
+    write_u32_le(f, 1);                   // Number of streams
+    write_u32_le(f, width * height * 3);  // Suggested buffer size
+    write_u32_le(f, width);
+    write_u32_le(f, height);
+    write_u32_le(f, 0);  // Reserved
+    write_u32_le(f, 0);  // Reserved
+    write_u32_le(f, 0);  // Reserved
+    write_u32_le(f, 0);  // Reserved
+
+    // 'strl' LIST (stream list)
+    fwrite("LIST", 4, 1, f);
+    write_u32_le(f, 4 + 8 + 56 + 8 + 40);
+    fwrite("strl", 4, 1, f);
+
+    // 'strh' chunk (stream header)
+    fwrite("strh", 4, 1, f);
+    write_u32_le(f, 56);
+    fwrite("vids", 4, 1, f);              // Stream type: video
+    fwrite("MJPG", 4, 1, f);              // Codec: Motion JPEG
+    write_u32_le(f, 0);                   // Flags
+    write_u16_le(f, 0);                   // Priority
+    write_u16_le(f, 0);                   // Language
+    write_u32_le(f, 0);                   // Initial frames
+    write_u32_le(f, 1);                   // Scale
+    write_u32_le(f, fps);                 // Rate
+    write_u32_le(f, 0);                   // Start
+    write_u32_le(f, num_images);          // Length
+    write_u32_le(f, width * height * 3);  // Suggested buffer size
+    write_u32_le(f, (uint32_t)-1);        // Quality
+    write_u32_le(f, 0);                   // Sample size
+    write_u16_le(f, 0);                   // rcFrame.left
+    write_u16_le(f, 0);                   // rcFrame.top
+    write_u16_le(f, 0);                   // rcFrame.right
+    write_u16_le(f, 0);                   // rcFrame.bottom
+
+    // 'strf' chunk (stream format: BITMAPINFOHEADER)
+    fwrite("strf", 4, 1, f);
+    write_u32_le(f, 40);
+    write_u32_le(f, 40);  // biSize
+    write_u32_le(f, width);
+    write_u32_le(f, height);
+    write_u16_le(f, 1);                   // biPlanes
+    write_u16_le(f, 24);                  // biBitCount
+    fwrite("MJPG", 4, 1, f);              // biCompression (FOURCC)
+    write_u32_le(f, width * height * 3);  // biSizeImage
+    write_u32_le(f, 0);                   // XPelsPerMeter
+    write_u32_le(f, 0);                   // YPelsPerMeter
+    write_u32_le(f, 0);                   // Colors used
+    write_u32_le(f, 0);                   // Colors important
+
+    // 'movi' LIST (video frames)
+    long movi_list_pos = ftell(f);
+    fwrite("LIST", 4, 1, f);
+    long movi_size_pos = ftell(f);
+    write_u32_le(f, 0);  // Placeholder for movi size
+    fwrite("movi", 4, 1, f);
+
+    avi_index_entry* index = (avi_index_entry*)malloc(sizeof(avi_index_entry) * num_images);
+    if (!index) {
+        fclose(f);
+        return -1;
+    }
+
+    // Encode and write each frame as JPEG
+    struct {
+        uint8_t* buf;
+        size_t size;
+    } jpeg_data;
+
+    for (int i = 0; i < num_images; i++) {
+        jpeg_data.buf  = NULL;
+        jpeg_data.size = 0;
+
+        // Callback function to collect JPEG data into memory
+        auto write_to_buf = [](void* context, void* data, int size) {
+            auto jd = (decltype(jpeg_data)*)context;
+            jd->buf = (uint8_t*)realloc(jd->buf, jd->size + size);
+            memcpy(jd->buf + jd->size, data, size);
+            jd->size += size;
+        };
+
+        // Encode to JPEG in memory
+        stbi_write_jpg_to_func(
+            write_to_buf,
+            &jpeg_data,
+            images[i].width,
+            images[i].height,
+            channels,
+            images[i].data,
+            quality);
+
+        // Write '00dc' chunk (video frame)
+        fwrite("00dc", 4, 1, f);
+        write_u32_le(f, jpeg_data.size);
+        index[i].offset = ftell(f) - 8;
+        index[i].size   = jpeg_data.size;
+        fwrite(jpeg_data.buf, 1, jpeg_data.size, f);
+
+        // Align to even byte size
+        if (jpeg_data.size % 2)
+            fputc(0, f);
+
+        free(jpeg_data.buf);
+    }
+
+    // Finalize 'movi' size
+    long cur_pos   = ftell(f);
+    long movi_size = cur_pos - movi_size_pos - 4;
+    fseek(f, movi_size_pos, SEEK_SET);
+    write_u32_le(f, movi_size);
+    fseek(f, cur_pos, SEEK_SET);
+
+    // Write 'idx1' index
+    fwrite("idx1", 4, 1, f);
+    write_u32_le(f, num_images * 16);
+    for (int i = 0; i < num_images; i++) {
+        fwrite("00dc", 4, 1, f);
+        write_u32_le(f, 0x10);
+        write_u32_le(f, index[i].offset);
+        write_u32_le(f, index[i].size);
+    }
+
+    // Finalize RIFF size
+    cur_pos        = ftell(f);
+    long file_size = cur_pos - riff_size_pos - 4;
+    fseek(f, riff_size_pos, SEEK_SET);
+    write_u32_le(f, file_size);
+    fseek(f, cur_pos, SEEK_SET);
+
+    fclose(f);
+    free(index);
+
+    return 0;
+}
+
+#endif  // __AVI_WRITER_H__

flux.hpp

@@ -5,6 +5,7 @@
 
 #include "ggml_extend.hpp"
 #include "model.h"
+#include "rope.hpp"
 
 #define FLUX_GRAPH_SIZE 10240
 
@@ -35,8 +36,8 @@ namespace Flux {
         int64_t hidden_size;
         float eps;
 
-        void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
-            ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "scale") != tensor_types.end()) ? tensor_types[prefix + "scale"] : GGML_TYPE_F32;
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+            ggml_type wtype = GGML_TYPE_F32;
             params["scale"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
         }
 
@@ -113,6 +114,7 @@ namespace Flux {
     }
 
     __STATIC_INLINE__ struct ggml_tensor* attention(struct ggml_context* ctx,
+                                                    ggml_backend_t backend,
                                                     struct ggml_tensor* q,
                                                     struct ggml_tensor* k,
                                                     struct ggml_tensor* v,
@@ -125,7 +127,7 @@ namespace Flux {
         q = apply_rope(ctx, q, pe);  // [N*n_head, L, d_head]
         k = apply_rope(ctx, k, pe);  // [N*n_head, L, d_head]
 
-        auto x = ggml_nn_attention_ext(ctx, q, k, v, v->ne[1], mask, false, true, flash_attn);  // [N, L, n_head*d_head]
+        auto x = ggml_nn_attention_ext(ctx, backend, q, k, v, v->ne[1], mask, false, true, flash_attn);  // [N, L, n_head*d_head]
         return x;
     }
 
@@ -168,13 +170,17 @@ namespace Flux {
             return x;
         }
 
-        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* pe, struct ggml_tensor* mask) {
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* pe,
+                                    struct ggml_tensor* mask) {
             // x: [N, n_token, dim]
             // pe: [n_token, d_head/2, 2, 2]
             // return [N, n_token, dim]
-            auto qkv = pre_attention(ctx, x);                                         // q,k,v: [N, n_token, n_head, d_head]
-            x        = attention(ctx, qkv[0], qkv[1], qkv[2], pe, mask, flash_attn);  // [N, n_token, dim]
-            x        = post_attention(ctx, x);                                        // [N, n_token, dim]
+            auto qkv = pre_attention(ctx, x);                                                  // q,k,v: [N, n_token, n_head, d_head]
+            x        = attention(ctx, backend, qkv[0], qkv[1], qkv[2], pe, mask, flash_attn);  // [N, n_token, dim]
+            x        = post_attention(ctx, x);                                                 // [N, n_token, dim]
             return x;
         }
     };
@@ -298,6 +304,7 @@ namespace Flux {
         }
 
         std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                    ggml_backend_t backend,
                                                                     struct ggml_tensor* img,
                                                                     struct ggml_tensor* txt,
                                                                     struct ggml_tensor* vec,
@@ -361,8 +368,8 @@ namespace Flux {
             auto k = ggml_concat(ctx, txt_k, img_k, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
             auto v = ggml_concat(ctx, txt_v, img_v, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
 
-            auto attn         = attention(ctx, q, k, v, pe, mask, flash_attn);        // [N, n_txt_token + n_img_token, n_head*d_head]
-            attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));  // [n_txt_token + n_img_token, N, hidden_size]
+            auto attn         = attention(ctx, backend, q, k, v, pe, mask, flash_attn);  // [N, n_txt_token + n_img_token, n_head*d_head]
+            attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));     // [n_txt_token + n_img_token, N, hidden_size]
             auto txt_attn_out = ggml_view_3d(ctx,
                                              attn,
                                              attn->ne[0],
@@ -445,6 +452,7 @@ namespace Flux {
         }
 
         struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* vec,
                                     struct ggml_tensor* pe,
@@ -495,7 +503,7 @@ namespace Flux {
             auto v           = ggml_reshape_4d(ctx, qkv_vec[2], head_dim, num_heads, qkv_vec[2]->ne[1], qkv_vec[2]->ne[2]);  // [N, n_token, n_head, d_head]
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
-            auto attn        = attention(ctx, q, k, v, pe, mask, flash_attn);  // [N, n_token, hidden_size]
+            auto attn        = attention(ctx, backend, q, k, v, pe, mask, flash_attn);  // [N, n_token, hidden_size]
 
             auto attn_mlp = ggml_concat(ctx, attn, ggml_gelu_inplace(ctx, mlp), 0);  // [N, n_token, hidden_size + mlp_hidden_dim]
             auto output   = linear2->forward(ctx, attn_mlp);                         // [N, n_token, hidden_size]
@@ -610,179 +618,11 @@ namespace Flux {
     };
 
     struct Flux : public GGMLBlock {
-    public:
-        std::vector<float> linspace(float start, float end, int num) {
-            std::vector<float> result(num);
-            float step = (end - start) / (num - 1);
-            for (int i = 0; i < num; ++i) {
-                result[i] = start + i * step;
-            }
-            return result;
-        }
-
-        std::vector<std::vector<float>> transpose(const std::vector<std::vector<float>>& mat) {
-            int rows = mat.size();
-            int cols = mat[0].size();
-            std::vector<std::vector<float>> transposed(cols, std::vector<float>(rows));
-            for (int i = 0; i < rows; ++i) {
-                for (int j = 0; j < cols; ++j) {
-                    transposed[j][i] = mat[i][j];
-                }
-            }
-            return transposed;
-        }
-
-        std::vector<float> flatten(const std::vector<std::vector<float>>& vec) {
-            std::vector<float> flat_vec;
-            for (const auto& sub_vec : vec) {
-                flat_vec.insert(flat_vec.end(), sub_vec.begin(), sub_vec.end());
-            }
-            return flat_vec;
-        }
-
-        std::vector<std::vector<float>> rope(const std::vector<float>& pos, int dim, int theta) {
-            assert(dim % 2 == 0);
-            int half_dim = dim / 2;
-
-            std::vector<float> scale = linspace(0, (dim * 1.0f - 2) / dim, half_dim);
-
-            std::vector<float> omega(half_dim);
-            for (int i = 0; i < half_dim; ++i) {
-                omega[i] = 1.0 / std::pow(theta, scale[i]);
-            }
-
-            int pos_size = pos.size();
-            std::vector<std::vector<float>> out(pos_size, std::vector<float>(half_dim));
-            for (int i = 0; i < pos_size; ++i) {
-                for (int j = 0; j < half_dim; ++j) {
-                    out[i][j] = pos[i] * omega[j];
-                }
-            }
-
-            std::vector<std::vector<float>> result(pos_size, std::vector<float>(half_dim * 4));
-            for (int i = 0; i < pos_size; ++i) {
-                for (int j = 0; j < half_dim; ++j) {
-                    result[i][4 * j]     = std::cos(out[i][j]);
-                    result[i][4 * j + 1] = -std::sin(out[i][j]);
-                    result[i][4 * j + 2] = std::sin(out[i][j]);
-                    result[i][4 * j + 3] = std::cos(out[i][j]);
-                }
-            }
-
-            return result;
-        }
-
-        // Generate IDs for image patches and text
-        std::vector<std::vector<float>> gen_txt_ids(int bs, int context_len) {
-            return std::vector<std::vector<float>>(bs * context_len, std::vector<float>(3, 0.0));
-        }
-
-        std::vector<std::vector<float>> gen_img_ids(int h, int w, int patch_size, int bs, int index = 0, int h_offset = 0, int w_offset = 0) {
-            int h_len = (h + (patch_size / 2)) / patch_size;
-            int w_len = (w + (patch_size / 2)) / patch_size;
-
-            std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(3, 0.0));
-
-            std::vector<float> row_ids = linspace(h_offset, h_len - 1 + h_offset, h_len);
-            std::vector<float> col_ids = linspace(w_offset, w_len - 1 + w_offset, w_len);
-
-            for (int i = 0; i < h_len; ++i) {
-                for (int j = 0; j < w_len; ++j) {
-                    img_ids[i * w_len + j][0] = index;
-                    img_ids[i * w_len + j][1] = row_ids[i];
-                    img_ids[i * w_len + j][2] = col_ids[j];
-                }
-            }
-
-            std::vector<std::vector<float>> img_ids_repeated(bs * img_ids.size(), std::vector<float>(3));
-            for (int i = 0; i < bs; ++i) {
-                for (int j = 0; j < img_ids.size(); ++j) {
-                    img_ids_repeated[i * img_ids.size() + j] = img_ids[j];
-                }
-            }
-            return img_ids_repeated;
-        }
-
-        std::vector<std::vector<float>> concat_ids(const std::vector<std::vector<float>>& a,
-                                                   const std::vector<std::vector<float>>& b,
-                                                   int bs) {
-            size_t a_len = a.size() / bs;
-            size_t b_len = b.size() / bs;
-            std::vector<std::vector<float>> ids(a.size() + b.size(), std::vector<float>(3));
-            for (int i = 0; i < bs; ++i) {
-                for (int j = 0; j < a_len; ++j) {
-                    ids[i * (a_len + b_len) + j] = a[i * a_len + j];
-                }
-                for (int j = 0; j < b_len; ++j) {
-                    ids[i * (a_len + b_len) + a_len + j] = b[i * b_len + j];
-                }
-            }
-            return ids;
-        }
-
-        std::vector<std::vector<float>> gen_ids(int h, int w, int patch_size, int bs, int context_len, std::vector<ggml_tensor*> ref_latents) {
-            auto txt_ids = gen_txt_ids(bs, context_len);
-            auto img_ids = gen_img_ids(h, w, patch_size, bs);
-
-            auto ids               = concat_ids(txt_ids, img_ids, bs);
-            uint64_t curr_h_offset = 0;
-            uint64_t curr_w_offset = 0;
-            for (ggml_tensor* ref : ref_latents) {
-                uint64_t h_offset = 0;
-                uint64_t w_offset = 0;
-                if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) {
-                    w_offset = curr_w_offset;
-                } else {
-                    h_offset = curr_h_offset;
-                }
-
-                auto ref_ids = gen_img_ids(ref->ne[1], ref->ne[0], patch_size, bs, 1, h_offset, w_offset);
-                ids          = concat_ids(ids, ref_ids, bs);
-
-                curr_h_offset = std::max(curr_h_offset, ref->ne[1] + h_offset);
-                curr_w_offset = std::max(curr_w_offset, ref->ne[0] + w_offset);
-            }
-            return ids;
-        }
-
-        // Generate positional embeddings
-        std::vector<float> gen_pe(int h, int w, int patch_size, int bs, int context_len, std::vector<ggml_tensor*> ref_latents, int theta, const std::vector<int>& axes_dim) {
-            std::vector<std::vector<float>> ids       = gen_ids(h, w, patch_size, bs, context_len, ref_latents);
-            std::vector<std::vector<float>> trans_ids = transpose(ids);
-            size_t pos_len                            = ids.size();
-            int num_axes                              = axes_dim.size();
-            for (int i = 0; i < pos_len; i++) {
-                // std::cout << trans_ids[0][i] << " " << trans_ids[1][i] << " " << trans_ids[2][i] << std::endl;
-            }
-
-            int emb_dim = 0;
-            for (int d : axes_dim)
-                emb_dim += d / 2;
-
-            std::vector<std::vector<float>> emb(bs * pos_len, std::vector<float>(emb_dim * 2 * 2, 0.0));
-            int offset = 0;
-            for (int i = 0; i < num_axes; ++i) {
-                std::vector<std::vector<float>> rope_emb = rope(trans_ids[i], axes_dim[i], theta);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
-                for (int b = 0; b < bs; ++b) {
-                    for (int j = 0; j < pos_len; ++j) {
-                        for (int k = 0; k < rope_emb[0].size(); ++k) {
-                            emb[b * pos_len + j][offset + k] = rope_emb[j][k];
-                        }
-                    }
-                }
-                offset += rope_emb[0].size();
-            }
-
-            return flatten(emb);
-        }
-
     public:
         FluxParams params;
         Flux() {}
         Flux(FluxParams params)
             : params(params) {
-            int64_t pe_dim = params.hidden_size / params.num_heads;
-
             blocks["img_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, params.hidden_size, true));
             if (params.is_chroma) {
                 blocks["distilled_guidance_layer"] = std::shared_ptr<GGMLBlock>(new ChromaApproximator(params.in_channels, params.hidden_size));
@@ -866,6 +706,7 @@ namespace Flux {
         }
 
         struct ggml_tensor* forward_orig(struct ggml_context* ctx,
+                                         ggml_backend_t backend,
                                          struct ggml_tensor* img,
                                          struct ggml_tensor* txt,
                                          struct ggml_tensor* timesteps,
@@ -930,7 +771,7 @@ namespace Flux {
 
                 auto block = std::dynamic_pointer_cast<DoubleStreamBlock>(blocks["double_blocks." + std::to_string(i)]);
 
-                auto img_txt = block->forward(ctx, img, txt, vec, pe, txt_img_mask);
+                auto img_txt = block->forward(ctx, backend, img, txt, vec, pe, txt_img_mask);
                 img          = img_txt.first;   // [N, n_img_token, hidden_size]
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
             }
@@ -942,7 +783,7 @@ namespace Flux {
                 }
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
-                txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask);
+                txt_img = block->forward(ctx, backend, txt_img, vec, pe, txt_img_mask);
             }
 
             txt_img = ggml_cont(ctx, ggml_permute(ctx, txt_img, 0, 2, 1, 3));  // [n_txt_token + n_img_token, N, hidden_size]
@@ -975,6 +816,7 @@ namespace Flux {
         }
 
         struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* timestep,
                                     struct ggml_tensor* context,
@@ -1024,7 +866,7 @@ namespace Flux {
                 }
             }
 
-            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers);  // [N, num_tokens, C * patch_size * patch_size]
+            auto out = forward_orig(ctx, backend, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers);  // [N, num_tokens, C * patch_size * patch_size]
             if (out->ne[1] > img_tokens) {
                 out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));  // [num_tokens, N, C * patch_size * patch_size]
                 out = ggml_view_3d(ctx, out, out->ne[0], out->ne[1], img_tokens, out->nb[1], out->nb[2], 0);
@@ -1039,8 +881,6 @@ namespace Flux {
     };
 
     struct FluxRunner : public GGMLRunner {
-        static std::map<std::string, enum ggml_type> empty_tensor_types;
-
     public:
         FluxParams flux_params;
         Flux flux;
@@ -1050,12 +890,13 @@ namespace Flux {
         bool use_mask = false;
 
         FluxRunner(ggml_backend_t backend,
-                   std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
-                   const std::string prefix                            = "",
-                   SDVersion version                                   = VERSION_FLUX,
-                   bool flash_attn                                     = false,
-                   bool use_mask                                       = false)
-            : GGMLRunner(backend), use_mask(use_mask) {
+                   bool offload_params_to_cpu,
+                   const String2GGMLType& tensor_types = {},
+                   const std::string prefix            = "",
+                   SDVersion version                   = VERSION_FLUX,
+                   bool flash_attn                     = false,
+                   bool use_mask                       = false)
+            : GGMLRunner(backend, offload_params_to_cpu), use_mask(use_mask) {
             flux_params.flash_attn          = flash_attn;
             flux_params.guidance_embed      = false;
             flux_params.depth               = 0;
@@ -1065,7 +906,7 @@ namespace Flux {
             }
             for (auto pair : tensor_types) {
                 std::string tensor_name = pair.first;
-                if (tensor_name.find("model.diffusion_model.") == std::string::npos)
+                if (!starts_with(tensor_name, prefix))
                     continue;
                 if (tensor_name.find("guidance_in.in_layer.weight") != std::string::npos) {
                     // not schnell
@@ -1119,6 +960,7 @@ namespace Flux {
                                         struct ggml_tensor* y,
                                         struct ggml_tensor* guidance,
                                         std::vector<ggml_tensor*> ref_latents = {},
+                                        bool increase_ref_index               = false,
                                         std::vector<int> skip_layers          = {}) {
             GGML_ASSERT(x->ne[3] == 1);
             struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, FLUX_GRAPH_SIZE, false);
@@ -1152,7 +994,15 @@ namespace Flux {
                 ref_latents[i] = to_backend(ref_latents[i]);
             }
 
-            pe_vec      = flux.gen_pe(x->ne[1], x->ne[0], 2, x->ne[3], context->ne[1], ref_latents, flux_params.theta, flux_params.axes_dim);
+            pe_vec      = Rope::gen_flux_pe(x->ne[1],
+                                            x->ne[0],
+                                            2,
+                                            x->ne[3],
+                                            context->ne[1],
+                                            ref_latents,
+                                            increase_ref_index,
+                                            flux_params.theta,
+                                            flux_params.axes_dim);
             int pos_len = pe_vec.size() / flux_params.axes_dim_sum / 2;
             // LOG_DEBUG("pos_len %d", pos_len);
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len);
@@ -1162,6 +1012,7 @@ namespace Flux {
             set_backend_tensor_data(pe, pe_vec.data());
 
             struct ggml_tensor* out = flux.forward(compute_ctx,
+                                                   runtime_backend,
                                                    x,
                                                    timesteps,
                                                    context,
@@ -1186,6 +1037,7 @@ namespace Flux {
                      struct ggml_tensor* y,
                      struct ggml_tensor* guidance,
                      std::vector<ggml_tensor*> ref_latents = {},
+                     bool increase_ref_index               = false,
                      struct ggml_tensor** output           = NULL,
                      struct ggml_context* output_ctx       = NULL,
                      std::vector<int> skip_layers          = std::vector<int>()) {
@@ -1195,7 +1047,7 @@ namespace Flux {
             // y: [N, adm_in_channels] or [1, adm_in_channels]
             // guidance: [N, ]
             auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, skip_layers);
+                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, increase_ref_index, skip_layers);
             };
 
             GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
@@ -1235,7 +1087,7 @@ namespace Flux {
                 struct ggml_tensor* out = NULL;
 
                 int t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, NULL, y, guidance, {}, &out, work_ctx);
+                compute(8, x, timesteps, context, NULL, y, guidance, {}, false, &out, work_ctx);
                 int t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -1247,7 +1099,7 @@ namespace Flux {
             // ggml_backend_t backend    = ggml_backend_cuda_init(0);
             ggml_backend_t backend           = ggml_backend_cpu_init();
             ggml_type model_data_type        = GGML_TYPE_Q8_0;
-            std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend));
+            std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend, false));
             {
                 LOG_INFO("loading from '%s'", file_path.c_str());
 
@@ -1261,7 +1113,7 @@ namespace Flux {
                     return;
                 }
 
-                bool success = model_loader.load_tensors(tensors, backend);
+                bool success = model_loader.load_tensors(tensors);
 
                 if (!success) {
                     LOG_ERROR("load tensors from model loader failed");

format-code.sh

@@ -1,2 +1,5 @@
-clang-format -style=file -i *.cpp *.h *.hpp
-clang-format -style=file -i examples/cli/*.cpp
+for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/cli/*.h; do
+  [[ "$f" == vocab* ]] && continue
+  echo "formatting '$f'"
+  clang-format -style=file -i "$f"
+done

gguf_reader.hpp

@@ -0,0 +1,231 @@
+#ifndef __GGUF_READER_HPP__
+#define __GGUF_READER_HPP__
+
+#include <cstdint>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include "ggml.h"
+#include "util.h"
+
+struct GGUFTensorInfo {
+    std::string name;
+    ggml_type type;
+    std::vector<int64_t> shape;
+    size_t offset;
+};
+
+enum class GGUFMetadataType : uint32_t {
+    UINT8   = 0,
+    INT8    = 1,
+    UINT16  = 2,
+    INT16   = 3,
+    UINT32  = 4,
+    INT32   = 5,
+    FLOAT32 = 6,
+    BOOL    = 7,
+    STRING  = 8,
+    ARRAY   = 9,
+    UINT64  = 10,
+    INT64   = 11,
+    FLOAT64 = 12,
+};
+
+class GGUFReader {
+private:
+    std::vector<GGUFTensorInfo> tensors_;
+    size_t data_offset_;
+    size_t alignment_ = 32;  // default alignment is 32
+
+    template <typename T>
+    bool safe_read(std::ifstream& fin, T& value) {
+        fin.read(reinterpret_cast<char*>(&value), sizeof(T));
+        return fin.good();
+    }
+
+    bool safe_read(std::ifstream& fin, char* buffer, size_t size) {
+        fin.read(buffer, size);
+        return fin.good();
+    }
+
+    bool safe_seek(std::ifstream& fin, std::streamoff offset, std::ios::seekdir dir) {
+        fin.seekg(offset, dir);
+        return fin.good();
+    }
+
+    bool read_metadata(std::ifstream& fin) {
+        uint64_t key_len = 0;
+        if (!safe_read(fin, key_len))
+            return false;
+
+        std::string key(key_len, '\0');
+        if (!safe_read(fin, (char*)key.data(), key_len))
+            return false;
+
+        uint32_t type = 0;
+        if (!safe_read(fin, type))
+            return false;
+
+        if (key == "general.alignment") {
+            uint32_t align_val = 0;
+            if (!safe_read(fin, align_val))
+                return false;
+
+            if (align_val != 0 && (align_val & (align_val - 1)) == 0) {
+                alignment_ = align_val;
+                LOG_DEBUG("Found alignment: %zu", alignment_);
+            } else {
+                LOG_ERROR("Invalid alignment value %u, fallback to default %zu", align_val, alignment_);
+            }
+            return true;
+        }
+
+        switch (static_cast<GGUFMetadataType>(type)) {
+            case GGUFMetadataType::UINT8:
+            case GGUFMetadataType::INT8:
+            case GGUFMetadataType::BOOL:
+                return safe_seek(fin, 1, std::ios::cur);
+
+            case GGUFMetadataType::UINT16:
+            case GGUFMetadataType::INT16:
+                return safe_seek(fin, 2, std::ios::cur);
+
+            case GGUFMetadataType::UINT32:
+            case GGUFMetadataType::INT32:
+            case GGUFMetadataType::FLOAT32:
+                return safe_seek(fin, 4, std::ios::cur);
+
+            case GGUFMetadataType::UINT64:
+            case GGUFMetadataType::INT64:
+            case GGUFMetadataType::FLOAT64:
+                return safe_seek(fin, 8, std::ios::cur);
+
+            case GGUFMetadataType::STRING: {
+                uint64_t len = 0;
+                if (!safe_read(fin, len))
+                    return false;
+                return safe_seek(fin, len, std::ios::cur);
+            }
+
+            case GGUFMetadataType::ARRAY: {
+                uint32_t elem_type = 0;
+                uint64_t len       = 0;
+                if (!safe_read(fin, elem_type))
+                    return false;
+                if (!safe_read(fin, len))
+                    return false;
+
+                for (uint64_t i = 0; i < len; i++) {
+                    if (!read_metadata(fin))
+                        return false;
+                }
+                return true;
+            }
+
+            default:
+                LOG_ERROR("Unknown metadata type=%u", type);
+                return false;
+        }
+    }
+
+    GGUFTensorInfo read_tensor_info(std::ifstream& fin) {
+        GGUFTensorInfo info;
+
+        uint64_t name_len;
+        if (!safe_read(fin, name_len))
+            throw std::runtime_error("read tensor name length failed");
+
+        info.name.resize(name_len);
+        if (!safe_read(fin, (char*)info.name.data(), name_len))
+            throw std::runtime_error("read tensor name failed");
+
+        uint32_t n_dims;
+        if (!safe_read(fin, n_dims))
+            throw std::runtime_error("read tensor dims failed");
+
+        info.shape.resize(n_dims);
+        for (uint32_t i = 0; i < n_dims; i++) {
+            if (!safe_read(fin, info.shape[i]))
+                throw std::runtime_error("read tensor shape failed");
+        }
+
+        if (n_dims > GGML_MAX_DIMS) {
+            for (int i = GGML_MAX_DIMS; i < n_dims; i++) {
+                info.shape[GGML_MAX_DIMS - 1] *= info.shape[i];  // stack to last dim;
+            }
+            info.shape.resize(GGML_MAX_DIMS);
+            n_dims = GGML_MAX_DIMS;
+        }
+
+        uint32_t type;
+        if (!safe_read(fin, type))
+            throw std::runtime_error("read tensor type failed");
+        info.type = static_cast<ggml_type>(type);
+
+        if (!safe_read(fin, info.offset))
+            throw std::runtime_error("read tensor offset failed");
+
+        return info;
+    }
+
+public:
+    bool load(const std::string& file_path) {
+        std::ifstream fin(file_path, std::ios::binary);
+        if (!fin) {
+            LOG_ERROR("failed to open '%s'", file_path.c_str());
+            return false;
+        }
+
+        // --- Header ---
+        char magic[4];
+        if (!safe_read(fin, magic, 4) || strncmp(magic, "GGUF", 4) != 0) {
+            LOG_ERROR("not a valid GGUF file");
+            return false;
+        }
+
+        uint32_t version;
+        if (!safe_read(fin, version))
+            return false;
+
+        uint64_t tensor_count, metadata_kv_count;
+        if (!safe_read(fin, tensor_count))
+            return false;
+        if (!safe_read(fin, metadata_kv_count))
+            return false;
+
+        LOG_DEBUG("GGUF v%u, tensor_count=%llu, metadata_kv_count=%llu",
+                  version, (unsigned long long)tensor_count, (unsigned long long)metadata_kv_count);
+
+        // --- Read Metadata ---
+        for (uint64_t i = 0; i < metadata_kv_count; i++) {
+            if (!read_metadata(fin)) {
+                LOG_ERROR("read meta data failed");
+                return false;
+            }
+        }
+
+        // --- Tensor Infos ---
+        tensors_.clear();
+        try {
+            for (uint64_t i = 0; i < tensor_count; i++) {
+                tensors_.push_back(read_tensor_info(fin));
+            }
+        } catch (const std::runtime_error& e) {
+            LOG_ERROR("%s", e.what());
+            return false;
+        }
+
+        data_offset_ = static_cast<size_t>(fin.tellg());
+        if ((data_offset_ % alignment_) != 0) {
+            data_offset_ = ((data_offset_ + alignment_ - 1) / alignment_) * alignment_;
+        }
+        fin.close();
+        return true;
+    }
+
+    const std::vector<GGUFTensorInfo>& tensors() const { return tensors_; }
+    size_t data_offset() const { return data_offset_; }
+};
+
+#endif  // __GGUF_READER_HPP__

lora.hpp

@@ -1,6 +1,7 @@
 #ifndef __LORA_HPP__
 #define __LORA_HPP__
 
+#include <mutex>
 #include "ggml_extend.hpp"
 
 #define LORA_GRAPH_BASE_SIZE 10240
@@ -58,6 +59,7 @@ struct LoraModel : public GGMLRunner {
         {"x_block.attn.proj", "attn.to_out.0"},
         {"x_block.attn2.proj", "attn2.to_out.0"},
         // flux
+        {"img_in", "x_embedder"},
         // singlestream
         {"linear2", "proj_out"},
         {"modulation.lin", "norm.linear"},
@@ -92,6 +94,7 @@ struct LoraModel : public GGMLRunner {
 
     float multiplier = 1.0f;
     std::map<std::string, struct ggml_tensor*> lora_tensors;
+    std::map<ggml_tensor*, ggml_tensor*> original_tensor_to_final_tensor;
     std::string file_path;
     ModelLoader model_loader;
     bool load_failed                = false;
@@ -103,7 +106,7 @@ struct LoraModel : public GGMLRunner {
     LoraModel(ggml_backend_t backend,
               const std::string& file_path = "",
               const std::string prefix     = "")
-        : file_path(file_path), GGMLRunner(backend) {
+        : file_path(file_path), GGMLRunner(backend, false) {
         if (!model_loader.init_from_file(file_path, prefix)) {
             load_failed = true;
         }
@@ -113,7 +116,7 @@ struct LoraModel : public GGMLRunner {
         return "lora";
     }
 
-    bool load_from_file(bool filter_tensor = false) {
+    bool load_from_file(bool filter_tensor = false, int n_threads = 0) {
         LOG_INFO("loading LoRA from '%s'", file_path.c_str());
 
         if (load_failed) {
@@ -121,41 +124,53 @@ struct LoraModel : public GGMLRunner {
             return false;
         }
 
+        std::unordered_map<std::string, TensorStorage> tensors_to_create;
+        std::mutex lora_mutex;
         bool dry_run          = true;
         auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
-            const std::string& name = tensor_storage.name;
+            if (dry_run) {
+                const std::string& name = tensor_storage.name;
 
-            if (filter_tensor && !contains(name, "lora")) {
-                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
-                return true;
-            }
-            // LOG_INFO("%s", name.c_str());
-            for (int i = 0; i < LORA_TYPE_COUNT; i++) {
-                if (name.find(type_fingerprints[i]) != std::string::npos) {
-                    type = (lora_t)i;
-                    break;
+                if (filter_tensor && !contains(name, "lora")) {
+                    return true;
+                }
+
+                {
+                    std::lock_guard<std::mutex> lock(lora_mutex);
+                    for (int i = 0; i < LORA_TYPE_COUNT; i++) {
+                        if (name.find(type_fingerprints[i]) != std::string::npos) {
+                            type = (lora_t)i;
+                            break;
+                        }
+                    }
+                    tensors_to_create[name] = tensor_storage;
+                }
+            } else {
+                const std::string& name = tensor_storage.name;
+                auto iter               = lora_tensors.find(name);
+                if (iter != lora_tensors.end()) {
+                    *dst_tensor = iter->second;
                 }
             }
-
-            if (dry_run) {
-                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
-                                                           tensor_storage.type,
-                                                           tensor_storage.n_dims,
-                                                           tensor_storage.ne);
-                lora_tensors[name]       = real;
-            } else {
-                auto real   = lora_tensors[name];
-                *dst_tensor = real;
-            }
-
             return true;
         };
 
-        model_loader.load_tensors(on_new_tensor_cb, backend);
+        model_loader.load_tensors(on_new_tensor_cb, n_threads);
+
+        for (const auto& pair : tensors_to_create) {
+            const auto& name         = pair.first;
+            const auto& ts           = pair.second;
+            struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                       ts.type,
+                                                       ts.n_dims,
+                                                       ts.ne);
+            lora_tensors[name]       = real;
+        }
+
         alloc_params_buffer();
-        // exit(0);
+
         dry_run = false;
-        model_loader.load_tensors(on_new_tensor_cb, backend);
+        model_loader.load_tensors(on_new_tensor_cb, n_threads);
 
         LOG_DEBUG("lora type: \"%s\"/\"%s\"", lora_downs[type].c_str(), lora_ups[type].c_str());
 
@@ -167,6 +182,7 @@ struct LoraModel : public GGMLRunner {
         auto out = ggml_reshape_1d(ctx, a, ggml_nelements(a));
         out      = ggml_get_rows(ctx, out, zero_index);
         out      = ggml_reshape(ctx, out, a);
+        // auto out = ggml_cast(ctx, a, GGML_TYPE_F32);
         return out;
     }
 
@@ -245,14 +261,22 @@ struct LoraModel : public GGMLRunner {
         set_backend_tensor_data(zero_index, zero_index_vec.data());
         ggml_build_forward_expand(gf, zero_index);
 
+        original_tensor_to_final_tensor.clear();
+
         std::set<std::string> applied_lora_tensors;
         for (auto it : model_tensors) {
-            std::string k_tensor       = it.first;
-            struct ggml_tensor* weight = model_tensors[it.first];
+            std::string model_tensor_name    = it.first;
+            struct ggml_tensor* model_tensor = model_tensors[it.first];
 
-            std::vector<std::string> keys = to_lora_keys(k_tensor, version);
-            if (keys.size() == 0)
-                continue;
+            std::vector<std::string> keys = to_lora_keys(model_tensor_name, version);
+            bool is_bias                  = ends_with(model_tensor_name, ".bias");
+            if (keys.size() == 0) {
+                if (is_bias) {
+                    keys.push_back(model_tensor_name.substr(0, model_tensor_name.size() - 5));  // remove .bias
+                } else {
+                    continue;
+                }
+            }
 
             for (auto& key : keys) {
                 bool is_qkv_split = starts_with(key, "SPLIT|");
@@ -265,8 +289,22 @@ struct LoraModel : public GGMLRunner {
                 }
                 struct ggml_tensor* updown = NULL;
                 float scale_value          = 1.0f;
-                std::string fk             = lora_pre[type] + key;
-                if (lora_tensors.find(fk + ".hada_w1_a") != lora_tensors.end()) {
+                std::string full_key       = lora_pre[type] + key;
+                if (is_bias) {
+                    if (lora_tensors.find(full_key + ".diff_b") != lora_tensors.end()) {
+                        std::string diff_name = full_key + ".diff_b";
+                        ggml_tensor* diff     = lora_tensors[diff_name];
+                        updown                = to_f32(compute_ctx, diff);
+                        applied_lora_tensors.insert(diff_name);
+                    } else {
+                        continue;
+                    }
+                } else if (lora_tensors.find(full_key + ".diff") != lora_tensors.end()) {
+                    std::string diff_name = full_key + ".diff";
+                    ggml_tensor* diff     = lora_tensors[diff_name];
+                    updown                = to_f32(compute_ctx, diff);
+                    applied_lora_tensors.insert(diff_name);
+                } else if (lora_tensors.find(full_key + ".hada_w1_a") != lora_tensors.end()) {
                     // LoHa mode
 
                     // TODO: split qkv convention for LoHas (is it ever used?)
@@ -292,9 +330,9 @@ struct LoraModel : public GGMLRunner {
                     std::string hada_2_down_name = "";
                     std::string hada_2_up_name   = "";
 
-                    hada_1_down_name = fk + ".hada_w1_b";
-                    hada_1_up_name   = fk + ".hada_w1_a";
-                    hada_1_mid_name  = fk + ".hada_t1";
+                    hada_1_down_name = full_key + ".hada_w1_b";
+                    hada_1_up_name   = full_key + ".hada_w1_a";
+                    hada_1_mid_name  = full_key + ".hada_t1";
                     if (lora_tensors.find(hada_1_down_name) != lora_tensors.end()) {
                         hada_1_down = to_f32(compute_ctx, lora_tensors[hada_1_down_name]);
                     }
@@ -307,9 +345,9 @@ struct LoraModel : public GGMLRunner {
                         hada_1_up = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, hada_1_up));
                     }
 
-                    hada_2_down_name = fk + ".hada_w2_b";
-                    hada_2_up_name   = fk + ".hada_w2_a";
-                    hada_2_mid_name  = fk + ".hada_t2";
+                    hada_2_down_name = full_key + ".hada_w2_b";
+                    hada_2_up_name   = full_key + ".hada_w2_a";
+                    hada_2_mid_name  = full_key + ".hada_t2";
                     if (lora_tensors.find(hada_2_down_name) != lora_tensors.end()) {
                         hada_2_down = to_f32(compute_ctx, lora_tensors[hada_2_down_name]);
                     }
@@ -322,7 +360,7 @@ struct LoraModel : public GGMLRunner {
                         hada_2_up = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, hada_2_up));
                     }
 
-                    alpha_name = fk + ".alpha";
+                    alpha_name = full_key + ".alpha";
 
                     applied_lora_tensors.insert(hada_1_down_name);
                     applied_lora_tensors.insert(hada_1_up_name);
@@ -345,7 +383,7 @@ struct LoraModel : public GGMLRunner {
                         float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
                         scale_value = alpha / rank;
                     }
-                } else if (lora_tensors.find(fk + ".lokr_w1") != lora_tensors.end() || lora_tensors.find(fk + ".lokr_w1_a") != lora_tensors.end()) {
+                } else if (lora_tensors.find(full_key + ".lokr_w1") != lora_tensors.end() || lora_tensors.find(full_key + ".lokr_w1_a") != lora_tensors.end()) {
                     // LoKr mode
 
                     // TODO: split qkv convention for LoKrs (is it ever used?)
@@ -354,7 +392,7 @@ struct LoraModel : public GGMLRunner {
                         break;
                     }
 
-                    std::string alpha_name = fk + ".alpha";
+                    std::string alpha_name = full_key + ".alpha";
 
                     ggml_tensor* lokr_w1 = NULL;
                     ggml_tensor* lokr_w2 = NULL;
@@ -362,8 +400,8 @@ struct LoraModel : public GGMLRunner {
                     std::string lokr_w1_name = "";
                     std::string lokr_w2_name = "";
 
-                    lokr_w1_name = fk + ".lokr_w1";
-                    lokr_w2_name = fk + ".lokr_w2";
+                    lokr_w1_name = full_key + ".lokr_w1";
+                    lokr_w2_name = full_key + ".lokr_w2";
 
                     if (lora_tensors.find(lokr_w1_name) != lora_tensors.end()) {
                         lokr_w1 = to_f32(compute_ctx, lora_tensors[lokr_w1_name]);
@@ -435,29 +473,29 @@ struct LoraModel : public GGMLRunner {
 
                     if (is_qkv_split) {
                         std::string suffix  = "";
-                        auto split_q_d_name = fk + "q" + suffix + lora_downs[type] + ".weight";
+                        auto split_q_d_name = full_key + "q" + suffix + lora_downs[type] + ".weight";
 
                         if (lora_tensors.find(split_q_d_name) == lora_tensors.end()) {
                             suffix         = "_proj";
-                            split_q_d_name = fk + "q" + suffix + lora_downs[type] + ".weight";
+                            split_q_d_name = full_key + "q" + suffix + lora_downs[type] + ".weight";
                         }
                         if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
                             // print_ggml_tensor(it.second, true);  //[3072, 21504, 1, 1]
                             // find qkv and mlp up parts in LoRA model
-                            auto split_k_d_name = fk + "k" + suffix + lora_downs[type] + ".weight";
-                            auto split_v_d_name = fk + "v" + suffix + lora_downs[type] + ".weight";
+                            auto split_k_d_name = full_key + "k" + suffix + lora_downs[type] + ".weight";
+                            auto split_v_d_name = full_key + "v" + suffix + lora_downs[type] + ".weight";
 
-                            auto split_q_u_name = fk + "q" + suffix + lora_ups[type] + ".weight";
-                            auto split_k_u_name = fk + "k" + suffix + lora_ups[type] + ".weight";
-                            auto split_v_u_name = fk + "v" + suffix + lora_ups[type] + ".weight";
+                            auto split_q_u_name = full_key + "q" + suffix + lora_ups[type] + ".weight";
+                            auto split_k_u_name = full_key + "k" + suffix + lora_ups[type] + ".weight";
+                            auto split_v_u_name = full_key + "v" + suffix + lora_ups[type] + ".weight";
 
-                            auto split_q_scale_name = fk + "q" + suffix + ".scale";
-                            auto split_k_scale_name = fk + "k" + suffix + ".scale";
-                            auto split_v_scale_name = fk + "v" + suffix + ".scale";
+                            auto split_q_scale_name = full_key + "q" + suffix + ".scale";
+                            auto split_k_scale_name = full_key + "k" + suffix + ".scale";
+                            auto split_v_scale_name = full_key + "v" + suffix + ".scale";
 
-                            auto split_q_alpha_name = fk + "q" + suffix + ".alpha";
-                            auto split_k_alpha_name = fk + "k" + suffix + ".alpha";
-                            auto split_v_alpha_name = fk + "v" + suffix + ".alpha";
+                            auto split_q_alpha_name = full_key + "q" + suffix + ".alpha";
+                            auto split_k_alpha_name = full_key + "k" + suffix + ".alpha";
+                            auto split_v_alpha_name = full_key + "v" + suffix + ".alpha";
 
                             ggml_tensor* lora_q_down = NULL;
                             ggml_tensor* lora_q_up   = NULL;
@@ -571,29 +609,29 @@ struct LoraModel : public GGMLRunner {
                             applied_lora_tensors.insert(split_v_d_name);
                         }
                     } else if (is_qkvm_split) {
-                        auto split_q_d_name = fk + "attn.to_q" + lora_downs[type] + ".weight";
+                        auto split_q_d_name = full_key + "attn.to_q" + lora_downs[type] + ".weight";
                         if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
                             // print_ggml_tensor(it.second, true);  //[3072, 21504, 1, 1]
                             // find qkv and mlp up parts in LoRA model
-                            auto split_k_d_name = fk + "attn.to_k" + lora_downs[type] + ".weight";
-                            auto split_v_d_name = fk + "attn.to_v" + lora_downs[type] + ".weight";
+                            auto split_k_d_name = full_key + "attn.to_k" + lora_downs[type] + ".weight";
+                            auto split_v_d_name = full_key + "attn.to_v" + lora_downs[type] + ".weight";
 
-                            auto split_q_u_name = fk + "attn.to_q" + lora_ups[type] + ".weight";
-                            auto split_k_u_name = fk + "attn.to_k" + lora_ups[type] + ".weight";
-                            auto split_v_u_name = fk + "attn.to_v" + lora_ups[type] + ".weight";
+                            auto split_q_u_name = full_key + "attn.to_q" + lora_ups[type] + ".weight";
+                            auto split_k_u_name = full_key + "attn.to_k" + lora_ups[type] + ".weight";
+                            auto split_v_u_name = full_key + "attn.to_v" + lora_ups[type] + ".weight";
 
-                            auto split_m_d_name = fk + "proj_mlp" + lora_downs[type] + ".weight";
-                            auto split_m_u_name = fk + "proj_mlp" + lora_ups[type] + ".weight";
+                            auto split_m_d_name = full_key + "proj_mlp" + lora_downs[type] + ".weight";
+                            auto split_m_u_name = full_key + "proj_mlp" + lora_ups[type] + ".weight";
 
-                            auto split_q_scale_name = fk + "attn.to_q" + ".scale";
-                            auto split_k_scale_name = fk + "attn.to_k" + ".scale";
-                            auto split_v_scale_name = fk + "attn.to_v" + ".scale";
-                            auto split_m_scale_name = fk + "proj_mlp" + ".scale";
+                            auto split_q_scale_name = full_key + "attn.to_q" + ".scale";
+                            auto split_k_scale_name = full_key + "attn.to_k" + ".scale";
+                            auto split_v_scale_name = full_key + "attn.to_v" + ".scale";
+                            auto split_m_scale_name = full_key + "proj_mlp" + ".scale";
 
-                            auto split_q_alpha_name = fk + "attn.to_q" + ".alpha";
-                            auto split_k_alpha_name = fk + "attn.to_k" + ".alpha";
-                            auto split_v_alpha_name = fk + "attn.to_v" + ".alpha";
-                            auto split_m_alpha_name = fk + "proj_mlp" + ".alpha";
+                            auto split_q_alpha_name = full_key + "attn.to_q" + ".alpha";
+                            auto split_k_alpha_name = full_key + "attn.to_k" + ".alpha";
+                            auto split_v_alpha_name = full_key + "attn.to_v" + ".alpha";
+                            auto split_m_alpha_name = full_key + "proj_mlp" + ".alpha";
 
                             ggml_tensor* lora_q_down = NULL;
                             ggml_tensor* lora_q_up   = NULL;
@@ -748,30 +786,27 @@ struct LoraModel : public GGMLRunner {
                             applied_lora_tensors.insert(split_m_d_name);
                         }
                     } else {
-                        lora_up_name   = fk + lora_ups[type] + ".weight";
-                        lora_down_name = fk + lora_downs[type] + ".weight";
-                        lora_mid_name  = fk + ".lora_mid.weight";
+                        lora_up_name   = full_key + lora_ups[type] + ".weight";
+                        lora_down_name = full_key + lora_downs[type] + ".weight";
+                        lora_mid_name  = full_key + ".lora_mid.weight";
 
-                        alpha_name = fk + ".alpha";
-                        scale_name = fk + ".scale";
+                        alpha_name = full_key + ".alpha";
+                        scale_name = full_key + ".scale";
 
                         if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
                             lora_up = to_f32(compute_ctx, lora_tensors[lora_up_name]);
+                            applied_lora_tensors.insert(lora_up_name);
                         }
 
                         if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
                             lora_down = to_f32(compute_ctx, lora_tensors[lora_down_name]);
+                            applied_lora_tensors.insert(lora_down_name);
                         }
 
                         if (lora_tensors.find(lora_mid_name) != lora_tensors.end()) {
                             lora_mid = to_f32(compute_ctx, lora_tensors[lora_mid_name]);
                             applied_lora_tensors.insert(lora_mid_name);
                         }
-
-                        applied_lora_tensors.insert(lora_up_name);
-                        applied_lora_tensors.insert(lora_down_name);
-                        applied_lora_tensors.insert(alpha_name);
-                        applied_lora_tensors.insert(scale_name);
                     }
 
                     if (lora_up == NULL || lora_down == NULL) {
@@ -782,29 +817,37 @@ struct LoraModel : public GGMLRunner {
                     int64_t rank = lora_down->ne[ggml_n_dims(lora_down) - 1];
                     if (lora_tensors.find(scale_name) != lora_tensors.end()) {
                         scale_value = ggml_backend_tensor_get_f32(lora_tensors[scale_name]);
+                        applied_lora_tensors.insert(scale_name);
                     } else if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
                         float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
                         scale_value = alpha / rank;
+                        // LOG_DEBUG("rank %s %ld %.2f %.2f", alpha_name.c_str(), rank, alpha, scale_value);
+                        applied_lora_tensors.insert(alpha_name);
                     }
 
                     updown = ggml_merge_lora(compute_ctx, lora_down, lora_up, lora_mid);
                 }
                 scale_value *= multiplier;
-                updown = ggml_reshape(compute_ctx, updown, weight);
-                GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(weight));
+                ggml_tensor* original_tensor = model_tensor;
+                if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+                    model_tensor = ggml_dup_tensor(compute_ctx, model_tensor);
+                    set_backend_tensor_data(model_tensor, original_tensor->data);
+                }
+                updown = ggml_reshape(compute_ctx, updown, model_tensor);
+                GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(model_tensor));
                 updown = ggml_scale_inplace(compute_ctx, updown, scale_value);
-                ggml_tensor* final_weight;
-                if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
-                    // final_weight = ggml_new_tensor(compute_ctx, GGML_TYPE_F32, ggml_n_dims(weight), weight->ne);
-                    // final_weight = ggml_cpy(compute_ctx, weight, final_weight);
-                    final_weight = to_f32(compute_ctx, weight);
-                    final_weight = ggml_add_inplace(compute_ctx, final_weight, updown);
-                    final_weight = ggml_cpy(compute_ctx, final_weight, weight);
+                ggml_tensor* final_tensor;
+                if (model_tensor->type != GGML_TYPE_F32 && model_tensor->type != GGML_TYPE_F16) {
+                    final_tensor = to_f32(compute_ctx, model_tensor);
+                    final_tensor = ggml_add_inplace(compute_ctx, final_tensor, updown);
+                    final_tensor = ggml_cpy(compute_ctx, final_tensor, model_tensor);
                 } else {
-                    final_weight = ggml_add_inplace(compute_ctx, weight, updown);
+                    final_tensor = ggml_add_inplace(compute_ctx, model_tensor, updown);
+                }
+                ggml_build_forward_expand(gf, final_tensor);
+                if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+                    original_tensor_to_final_tensor[original_tensor] = final_tensor;
                 }
-                // final_weight = ggml_add_inplace(compute_ctx, weight, updown);  // apply directly
-                ggml_build_forward_expand(gf, final_weight);
                 break;
             }
         }
@@ -825,10 +868,10 @@ struct LoraModel : public GGMLRunner {
          * this function is called once to calculate the required buffer size
          * and then again to actually generate a graph to be used */
         if (applied_lora_tensors_count != total_lora_tensors_count) {
-            LOG_WARN("Only (%lu / %lu) LoRA tensors have been applied",
+            LOG_WARN("Only (%lu / %lu) LoRA tensors will be applied",
                      applied_lora_tensors_count, total_lora_tensors_count);
         } else {
-            LOG_DEBUG("(%lu / %lu) LoRA tensors applied successfully",
+            LOG_DEBUG("(%lu / %lu) LoRA tensors will be applied",
                       applied_lora_tensors_count, total_lora_tensors_count);
         }
 
@@ -839,7 +882,15 @@ struct LoraModel : public GGMLRunner {
         auto get_graph = [&]() -> struct ggml_cgraph* {
             return build_lora_graph(model_tensors, version);
         };
-        GGMLRunner::compute(get_graph, n_threads, true);
+        GGMLRunner::compute(get_graph, n_threads, false);
+        for (auto item : original_tensor_to_final_tensor) {
+            ggml_tensor* original_tensor = item.first;
+            ggml_tensor* final_tensor    = item.second;
+
+            ggml_backend_tensor_copy(final_tensor, original_tensor);
+        }
+        original_tensor_to_final_tensor.clear();
+        GGMLRunner::free_compute_buffer();
     }
 };
 

ltxv.hpp

@@ -0,0 +1,74 @@
+#ifndef __LTXV_HPP__
+#define __LTXV_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+
+namespace LTXV {
+
+    class CausalConv3d : public GGMLBlock {
+    protected:
+        int time_kernel_size;
+
+    public:
+        CausalConv3d(int64_t in_channels,
+                     int64_t out_channels,
+                     int kernel_size        = 3,
+                     std::tuple<int> stride = {1, 1, 1},
+                     int dilation           = 1,
+                     bool bias              = true) {
+            time_kernel_size = kernel_size / 2;
+            blocks["conv"]   = std::shared_ptr<GGMLBlock>(new Conv3d(in_channels,
+                                                                     out_channels,
+                                                                     {kernel_size, kernel_size, kernel_size},
+                                                                     stride,
+                                                                     {0, kernel_size / 2, kernel_size / 2},
+                                                                     {dilation, 1, 1},
+                                                                     bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* x,
+                                    bool causal = true) {
+            // x: [N*IC, ID, IH, IW]
+            // result: [N*OC, OD, OH, OW]
+            auto conv = std::dynamic_pointer_cast<Conv3d>(blocks["conv"]);
+            if (causal) {
+                auto h               = ggml_cont(ctx, ggml_permute(ctx, x, 0, 1, 3, 2));                                                  // [ID, N*IC, IH, IW]
+                auto first_frame     = ggml_view_3d(ctx, h, h->ne[0], h->ne[1], h->ne[2], h->nb[1], h->nb[2], 0);                         // [N*IC, IH, IW]
+                first_frame          = ggml_reshape_4d(ctx, first_frame, first_frame->ne[0], first_frame->ne[1], 1, first_frame->ne[2]);  // [N*IC, 1, IH, IW]
+                auto first_frame_pad = first_frame;
+                for (int i = 1; i < time_kernel_size - 1; i++) {
+                    first_frame_pad = ggml_concat(ctx, first_frame_pad, first_frame, 2);
+                }
+                x = ggml_concat(ctx, first_frame_pad, x, 2);
+            } else {
+                auto h         = ggml_cont(ctx, ggml_permute(ctx, x, 0, 1, 3, 2));  // [ID, N*IC, IH, IW]
+                int64_t offset = h->nb[2] * h->ne[2];
+
+                auto first_frame     = ggml_view_3d(ctx, h, h->ne[0], h->ne[1], h->ne[2], h->nb[1], h->nb[2], 0);                         // [N*IC, IH, IW]
+                first_frame          = ggml_reshape_4d(ctx, first_frame, first_frame->ne[0], first_frame->ne[1], 1, first_frame->ne[2]);  // [N*IC, 1, IH, IW]
+                auto first_frame_pad = first_frame;
+                for (int i = 1; i < (time_kernel_size - 1) / 2; i++) {
+                    first_frame_pad = ggml_concat(ctx, first_frame_pad, first_frame, 2);
+                }
+
+                auto last_frame     = ggml_view_3d(ctx, h, h->ne[0], h->ne[1], h->ne[2], h->nb[1], h->nb[2], offset * (h->ne[3] - 1));  // [N*IC, IH, IW]
+                last_frame          = ggml_reshape_4d(ctx, last_frame, last_frame->ne[0], last_frame->ne[1], 1, last_frame->ne[2]);     // [N*IC, 1, IH, IW]
+                auto last_frame_pad = last_frame;
+                for (int i = 1; i < (time_kernel_size - 1) / 2; i++) {
+                    last_frame_pad = ggml_concat(ctx, last_frame_pad, last_frame, 2);
+                }
+
+                x = ggml_concat(ctx, first_frame_pad, x, 2);
+                x = ggml_concat(ctx, x, last_frame_pad, 2);
+            }
+
+            x = conv->forward(ctx, x);
+            return x;
+        }
+    };
+
+};
+
+#endif

mmdit.hpp

@@ -142,43 +142,21 @@ public:
     }
 };
 
-class RMSNorm : public UnaryBlock {
-protected:
-    int64_t hidden_size;
-    float eps;
-
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
-        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
-        params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
-    }
-
-public:
-    RMSNorm(int64_t hidden_size,
-            float eps = 1e-06f)
-        : hidden_size(hidden_size),
-          eps(eps) {}
-
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
-        struct ggml_tensor* w = params["weight"];
-        x                     = ggml_rms_norm(ctx, x, eps);
-        x                     = ggml_mul(ctx, x, w);
-        return x;
-    }
-};
-
 class SelfAttention : public GGMLBlock {
 public:
     int64_t num_heads;
     bool pre_only;
     std::string qk_norm;
+    bool flash_attn;
 
 public:
     SelfAttention(int64_t dim,
                   int64_t num_heads   = 8,
                   std::string qk_norm = "",
                   bool qkv_bias       = false,
-                  bool pre_only       = false)
-        : num_heads(num_heads), pre_only(pre_only), qk_norm(qk_norm) {
+                  bool pre_only       = false,
+                  bool flash_attn     = false)
+        : num_heads(num_heads), pre_only(pre_only), qk_norm(qk_norm), flash_attn(flash_attn) {
         int64_t d_head = dim / num_heads;
         blocks["qkv"]  = std::shared_ptr<GGMLBlock>(new Linear(dim, dim * 3, qkv_bias));
         if (!pre_only) {
@@ -226,10 +204,12 @@ public:
     }
 
     // x: [N, n_token, dim]
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x) {
         auto qkv = pre_attention(ctx, x);
-        x        = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
-        x        = post_attention(ctx, x);                                         // [N, n_token, dim]
+        x        = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, NULL, false, false, true);  // [N, n_token, dim]
+        x        = post_attention(ctx, x);                                                                            // [N, n_token, dim]
         return x;
     }
 };
@@ -254,6 +234,7 @@ public:
     int64_t num_heads;
     bool pre_only;
     bool self_attn;
+    bool flash_attn;
 
 public:
     DismantledBlock(int64_t hidden_size,
@@ -262,16 +243,17 @@ public:
                     std::string qk_norm = "",
                     bool qkv_bias       = false,
                     bool pre_only       = false,
-                    bool self_attn      = false)
+                    bool self_attn      = false,
+                    bool flash_attn     = false)
         : num_heads(num_heads), pre_only(pre_only), self_attn(self_attn) {
         // rmsnorm is always Flase
         // scale_mod_only is always Flase
         // swiglu is always Flase
         blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
-        blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only));
+        blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only, flash_attn));
 
         if (self_attn) {
-            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false));
+            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false, flash_attn));
         }
 
         if (!pre_only) {
@@ -439,7 +421,10 @@ public:
         return x;
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* c) {
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* c) {
         // x: [N, n_token, hidden_size]
         // c: [N, hidden_size]
         // return: [N, n_token, hidden_size]
@@ -454,8 +439,8 @@ public:
             auto qkv2          = std::get<1>(qkv_intermediates);
             auto intermediates = std::get<2>(qkv_intermediates);
 
-            auto attn_out  = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads);     // [N, n_token, dim]
-            auto attn2_out = ggml_nn_attention_ext(ctx, qkv2[0], qkv2[1], qkv2[2], num_heads);  // [N, n_token, dim]
+            auto attn_out  = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, NULL, false, false, flash_attn);     // [N, n_token, dim]
+            auto attn2_out = ggml_nn_attention_ext(ctx, backend, qkv2[0], qkv2[1], qkv2[2], num_heads, NULL, false, false, flash_attn);  // [N, n_token, dim]
             x              = post_attention_x(ctx,
                                               attn_out,
                                               attn2_out,
@@ -471,7 +456,7 @@ public:
             auto qkv               = qkv_intermediates.first;
             auto intermediates     = qkv_intermediates.second;
 
-            auto attn_out = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
+            auto attn_out = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, NULL, false, false, flash_attn);  // [N, n_token, dim]
             x             = post_attention(ctx,
                                            attn_out,
                                            intermediates[0],
@@ -486,6 +471,8 @@ public:
 
 __STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*>
 block_mixing(struct ggml_context* ctx,
+             ggml_backend_t backend,
+             bool flash_attn,
              struct ggml_tensor* context,
              struct ggml_tensor* x,
              struct ggml_tensor* c,
@@ -515,8 +502,8 @@ block_mixing(struct ggml_context* ctx,
         qkv.push_back(ggml_concat(ctx, context_qkv[i], x_qkv[i], 1));
     }
 
-    auto attn         = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], x_block->num_heads);  // [N, n_context + n_token, hidden_size]
-    attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                     // [n_context + n_token, N, hidden_size]
+    auto attn         = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], x_block->num_heads, NULL, false, false, flash_attn);  // [N, n_context + n_token, hidden_size]
+    attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                                                              // [n_context + n_token, N, hidden_size]
     auto context_attn = ggml_view_3d(ctx,
                                      attn,
                                      attn->ne[0],
@@ -549,7 +536,7 @@ block_mixing(struct ggml_context* ctx,
     }
 
     if (x_block->self_attn) {
-        auto attn2 = ggml_nn_attention_ext(ctx, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads);  // [N, n_token, hidden_size]
+        auto attn2 = ggml_nn_attention_ext(ctx, backend, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads);  // [N, n_token, hidden_size]
 
         x = x_block->post_attention_x(ctx,
                                       x_attn,
@@ -574,6 +561,8 @@ block_mixing(struct ggml_context* ctx,
 }
 
 struct JointBlock : public GGMLBlock {
+    bool flash_attn;
+
 public:
     JointBlock(int64_t hidden_size,
                int64_t num_heads,
@@ -581,19 +570,22 @@ public:
                std::string qk_norm = "",
                bool qkv_bias       = false,
                bool pre_only       = false,
-               bool self_attn_x    = false) {
-        blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only));
-        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x));
+               bool self_attn_x    = false,
+               bool flash_attn     = false)
+        : flash_attn(flash_attn) {
+        blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only, false, flash_attn));
+        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x, flash_attn));
     }
 
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                ggml_backend_t backend,
                                                                 struct ggml_tensor* context,
                                                                 struct ggml_tensor* x,
                                                                 struct ggml_tensor* c) {
         auto context_block = std::dynamic_pointer_cast<DismantledBlock>(blocks["context_block"]);
         auto x_block       = std::dynamic_pointer_cast<DismantledBlock>(blocks["x_block"]);
 
-        return block_mixing(ctx, context, x, c, context_block, x_block);
+        return block_mixing(ctx, backend, flash_attn, context, x, c, context_block, x_block);
     }
 };
 
@@ -651,14 +643,16 @@ protected:
     int64_t context_embedder_out_dim = 1536;
     int64_t hidden_size;
     std::string qk_norm;
+    bool flash_attn = false;
 
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
-        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "pos_embed") != tensor_types.end()) ? tensor_types[prefix + "pos_embed"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F32;
         params["pos_embed"]  = ggml_new_tensor_3d(ctx, wtype, hidden_size, num_patchs, 1);
     }
 
 public:
-    MMDiT(std::map<std::string, enum ggml_type>& tensor_types) {
+    MMDiT(bool flash_attn = false, const String2GGMLType& tensor_types = {})
+        : flash_attn(flash_attn) {
         // input_size is always None
         // learn_sigma is always False
         // register_length is alwalys 0
@@ -726,7 +720,8 @@ public:
                                                                                                     qk_norm,
                                                                                                     true,
                                                                                                     i == depth - 1,
-                                                                                                    i <= d_self));
+                                                                                                    i <= d_self,
+                                                                                                    flash_attn));
         }
 
         blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
@@ -795,6 +790,7 @@ public:
     }
 
     struct ggml_tensor* forward_core_with_concat(struct ggml_context* ctx,
+                                                 ggml_backend_t backend,
                                                  struct ggml_tensor* x,
                                                  struct ggml_tensor* c_mod,
                                                  struct ggml_tensor* context,
@@ -813,7 +809,7 @@ public:
 
             auto block = std::dynamic_pointer_cast<JointBlock>(blocks["joint_blocks." + std::to_string(i)]);
 
-            auto context_x = block->forward(ctx, context, x, c_mod);
+            auto context_x = block->forward(ctx, backend, context, x, c_mod);
             context        = context_x.first;
             x              = context_x.second;
         }
@@ -824,6 +820,7 @@ public:
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* t,
                                 struct ggml_tensor* y        = NULL,
@@ -859,7 +856,7 @@ public:
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
         }
 
-        x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
+        x = forward_core_with_concat(ctx, backend, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
         x = unpatchify(ctx, x, h, w);  // [N, C, H, W]
 
@@ -869,12 +866,12 @@ public:
 struct MMDiTRunner : public GGMLRunner {
     MMDiT mmdit;
 
-    static std::map<std::string, enum ggml_type> empty_tensor_types;
-
     MMDiTRunner(ggml_backend_t backend,
-                std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
-                const std::string prefix                            = "")
-        : GGMLRunner(backend), mmdit(tensor_types) {
+                bool offload_params_to_cpu,
+                bool flash_attn,
+                const String2GGMLType& tensor_types = {},
+                const std::string prefix            = "")
+        : GGMLRunner(backend, offload_params_to_cpu), mmdit(flash_attn, tensor_types) {
         mmdit.init(params_ctx, tensor_types, prefix);
     }
 
@@ -899,6 +896,7 @@ struct MMDiTRunner : public GGMLRunner {
         timesteps = to_backend(timesteps);
 
         struct ggml_tensor* out = mmdit.forward(compute_ctx,
+                                                runtime_backend,
                                                 x,
                                                 timesteps,
                                                 y,
@@ -972,7 +970,7 @@ struct MMDiTRunner : public GGMLRunner {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
         ggml_backend_t backend             = ggml_backend_cpu_init();
         ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend));
+        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend, false, false));
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 
@@ -986,7 +984,7 @@ struct MMDiTRunner : public GGMLRunner {
                 return;
             }
 
-            bool success = model_loader.load_tensors(tensors, backend);
+            bool success = model_loader.load_tensors(tensors);
 
             if (!success) {
                 LOG_ERROR("load tensors from model loader failed");

model.cpp

@@ -1,15 +1,22 @@
 #include <stdarg.h>
+#include <atomic>
+#include <chrono>
 #include <fstream>
+#include <functional>
+#include <mutex>
 #include <regex>
 #include <set>
 #include <string>
+#include <thread>
 #include <unordered_map>
 #include <vector>
 
+#include "gguf_reader.hpp"
 #include "model.h"
 #include "stable-diffusion.h"
 #include "util.h"
 #include "vocab.hpp"
+#include "vocab_umt5.hpp"
 
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -88,6 +95,7 @@ const char* unused_tensors[] = {
     "posterior_mean_coef1",
     "posterior_mean_coef2",
     "cond_stage_model.transformer.text_model.embeddings.position_ids",
+    "cond_stage_model.transformer.vision_model.embeddings.position_ids",
     "cond_stage_model.model.logit_scale",
     "cond_stage_model.model.text_projection",
     "conditioner.embedders.0.transformer.text_model.embeddings.position_ids",
@@ -104,7 +112,7 @@ const char* unused_tensors[] = {
 };
 
 bool is_unused_tensor(std::string name) {
-    for (int i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
+    for (size_t i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
         if (starts_with(name, unused_tensors[i])) {
             return true;
         }
@@ -141,6 +149,11 @@ std::unordered_map<std::string, std::string> open_clip_to_hk_clip_resblock = {
     {"mlp.c_proj.weight", "mlp.fc2.weight"},
 };
 
+std::unordered_map<std::string, std::string> cond_model_name_map = {
+    {"transformer.vision_model.pre_layrnorm.weight", "transformer.vision_model.pre_layernorm.weight"},
+    {"transformer.vision_model.pre_layrnorm.bias", "transformer.vision_model.pre_layernorm.bias"},
+};
+
 std::unordered_map<std::string, std::string> vae_decoder_name_map = {
     {"first_stage_model.decoder.mid.attn_1.to_k.bias", "first_stage_model.decoder.mid.attn_1.k.bias"},
     {"first_stage_model.decoder.mid.attn_1.to_k.weight", "first_stage_model.decoder.mid.attn_1.k.weight"},
@@ -179,7 +192,7 @@ std::unordered_map<std::string, std::string> pmid_v2_name_map = {
      "pmid.qformer_perceiver.token_proj.fc2.weight"},
 };
 
-std::string convert_open_clip_to_hf_clip(const std::string& name) {
+std::string convert_cond_model_name(const std::string& name) {
     std::string new_name = name;
     std::string prefix;
     if (contains(new_name, ".enc.")) {
@@ -268,6 +281,10 @@ std::string convert_open_clip_to_hf_clip(const std::string& name) {
         new_name = open_clip_to_hf_clip_model[new_name];
     }
 
+    if (cond_model_name_map.find(new_name) != cond_model_name_map.end()) {
+        new_name = cond_model_name_map[new_name];
+    }
+
     std::string open_clip_resblock_prefix = "model.transformer.resblocks.";
     std::string hf_clip_resblock_prefix   = "transformer.text_model.encoder.layers.";
 
@@ -563,7 +580,7 @@ std::string convert_tensor_name(std::string name) {
     // }
     std::string new_name = name;
     if (starts_with(name, "cond_stage_model.") || starts_with(name, "conditioner.embedders.") || starts_with(name, "text_encoders.") || ends_with(name, ".vision_model.visual_projection.weight")) {
-        new_name = convert_open_clip_to_hf_clip(name);
+        new_name = convert_cond_model_name(name);
     } else if (starts_with(name, "first_stage_model.decoder")) {
         new_name = convert_vae_decoder_name(name);
     } else if (starts_with(name, "pmid.qformer_perceiver")) {
@@ -592,9 +609,11 @@ std::string convert_tensor_name(std::string name) {
         } else {
             new_name = name;
         }
+    } else if (ends_with(name, ".diff") || ends_with(name, ".diff_b")) {
+        new_name = "lora." + name;
     } else if (contains(name, "lora_up") || contains(name, "lora_down") ||
                contains(name, "lora.up") || contains(name, "lora.down") ||
-               contains(name, "lora_linear")) {
+               contains(name, "lora_linear") || ends_with(name, ".alpha")) {
         size_t pos = new_name.find(".processor");
         if (pos != std::string::npos) {
             new_name.replace(pos, strlen(".processor"), "");
@@ -602,7 +621,11 @@ std::string convert_tensor_name(std::string name) {
         // if (starts_with(new_name, "transformer.transformer_blocks") || starts_with(new_name, "transformer.single_transformer_blocks")) {
         //     new_name = "model.diffusion_model." + new_name;
         // }
-        pos = new_name.rfind("lora");
+        if (ends_with(name, ".alpha")) {
+            pos = new_name.rfind("alpha");
+        } else {
+            pos = new_name.rfind("lora");
+        }
         if (pos != std::string::npos) {
             std::string name_without_network_parts = new_name.substr(0, pos - 1);
             std::string network_part               = new_name.substr(pos);
@@ -648,7 +671,7 @@ std::string convert_tensor_name(std::string name) {
     return new_name;
 }
 
-void add_preprocess_tensor_storage_types(std::map<std::string, enum ggml_type>& tensor_storages_types, std::string name, enum ggml_type type) {
+void add_preprocess_tensor_storage_types(String2GGMLType& tensor_storages_types, std::string name, enum ggml_type type) {
     std::string new_name = convert_tensor_name(name);
 
     if (new_name.find("cond_stage_model") != std::string::npos && ends_with(new_name, "attn.in_proj_weight")) {
@@ -684,6 +707,13 @@ void preprocess_tensor(TensorStorage tensor_storage,
         tensor_storage.unsqueeze();
     }
 
+    // wan vae
+    if (ends_with(new_name, "gamma")) {
+        tensor_storage.reverse_ne();
+        tensor_storage.n_dims = 1;
+        tensor_storage.reverse_ne();
+    }
+
     tensor_storage.name = new_name;
 
     if (new_name.find("cond_stage_model") != std::string::npos &&
@@ -1030,10 +1060,38 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
 
     gguf_context* ctx_gguf_ = NULL;
     ggml_context* ctx_meta_ = NULL;
-    ctx_gguf_               = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
+
+    ctx_gguf_ = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
     if (!ctx_gguf_) {
-        LOG_ERROR("failed to open '%s'", file_path.c_str());
-        return false;
+        LOG_ERROR("failed to open '%s' with gguf_init_from_file. Try to open it with GGUFReader.", file_path.c_str());
+        GGUFReader gguf_reader;
+        if (!gguf_reader.load(file_path)) {
+            LOG_ERROR("failed to open '%s' with GGUFReader.", file_path.c_str());
+            return false;
+        }
+
+        size_t data_offset = gguf_reader.data_offset();
+        for (const auto& gguf_tensor_info : gguf_reader.tensors()) {
+            std::string name = gguf_tensor_info.name;
+            if (!starts_with(name, prefix)) {
+                name = prefix + name;
+            }
+
+            TensorStorage tensor_storage(
+                name,
+                gguf_tensor_info.type,
+                gguf_tensor_info.shape.data(),
+                gguf_tensor_info.shape.size(),
+                file_index,
+                data_offset + gguf_tensor_info.offset);
+
+            // LOG_DEBUG("%s %s", name.c_str(), tensor_storage.to_string().c_str());
+
+            tensor_storages.push_back(tensor_storage);
+            add_preprocess_tensor_storage_types(tensor_storages_types, tensor_storage.name, tensor_storage.type);
+        }
+
+        return true;
     }
 
     int n_tensors = gguf_get_n_tensors(ctx_gguf_);
@@ -1047,7 +1105,11 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
 
         // LOG_DEBUG("%s", name.c_str());
 
-        TensorStorage tensor_storage(prefix + name, dummy->type, dummy->ne, ggml_n_dims(dummy), file_index, offset);
+        if (!starts_with(name, prefix)) {
+            name = prefix + name;
+        }
+
+        TensorStorage tensor_storage(name, dummy->type, dummy->ne, ggml_n_dims(dummy), file_index, offset);
 
         GGML_ASSERT(ggml_nbytes(dummy) == tensor_storage.nbytes());
 
@@ -1085,7 +1147,7 @@ ggml_type str_to_ggml_type(const std::string& dtype) {
 
 // https://huggingface.co/docs/safetensors/index
 bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const std::string& prefix) {
-    LOG_DEBUG("init from '%s'", file_path.c_str());
+    LOG_DEBUG("init from '%s', prefix = '%s'", file_path.c_str(), prefix.c_str());
     file_paths_.push_back(file_path);
     size_t file_index = file_paths_.size() - 1;
     std::ifstream file(file_path, std::ios::binary);
@@ -1150,6 +1212,10 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
         std::string dtype    = tensor_info["dtype"];
         nlohmann::json shape = tensor_info["shape"];
 
+        if (dtype == "U8") {
+            continue;
+        }
+
         size_t begin = tensor_info["data_offsets"][0].get<size_t>();
         size_t end   = tensor_info["data_offsets"][1].get<size_t>();
 
@@ -1171,12 +1237,11 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
         }
 
         if (n_dims == 5) {
-            if (ne[3] == 1 && ne[4] == 1) {
-                n_dims = 4;
-            } else {
-                LOG_ERROR("invalid tensor '%s'", name.c_str());
-                return false;
-            }
+            n_dims = 4;
+            ne[0]  = ne[0] * ne[1];
+            ne[1]  = ne[2];
+            ne[2]  = ne[3];
+            ne[3]  = ne[4];
         }
 
         // ggml_n_dims returns 1 for scalars
@@ -1184,7 +1249,11 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
             n_dims = 1;
         }
 
-        TensorStorage tensor_storage(prefix + name, type, ne, n_dims, file_index, ST_HEADER_SIZE_LEN + header_size_ + begin);
+        if (!starts_with(name, prefix)) {
+            name = prefix + name;
+        }
+
+        TensorStorage tensor_storage(name, type, ne, n_dims, file_index, ST_HEADER_SIZE_LEN + header_size_ + begin);
         tensor_storage.reverse_ne();
 
         size_t tensor_data_size = end - begin;
@@ -1569,7 +1638,11 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
                         reader.tensor_storage.file_index = file_index;
                         // if(strcmp(prefix.c_str(), "scarlett") == 0)
                         // printf(" ZIP got tensor %s \n ", reader.tensor_storage.name.c_str());
-                        reader.tensor_storage.name = prefix + reader.tensor_storage.name;
+                        std::string name = reader.tensor_storage.name;
+                        if (!starts_with(name, prefix)) {
+                            name = prefix + name;
+                        }
+                        reader.tensor_storage.name = name;
                         tensor_storages.push_back(reader.tensor_storage);
                         add_preprocess_tensor_storage_types(tensor_storages_types, reader.tensor_storage.name, reader.tensor_storage.type);
 
@@ -1641,12 +1714,14 @@ SDVersion ModelLoader::get_sd_version() {
     bool has_multiple_encoders = false;
     bool is_unet               = false;
 
-    bool is_xl   = false;
-    bool is_flux = false;
+    bool is_xl                       = false;
+    bool is_flux                     = false;
+    bool is_wan                      = false;
+    int64_t patch_embedding_channels = 0;
+    bool has_img_emb                 = false;
 
-#define found_family (is_xl || is_flux)
     for (auto& tensor_storage : tensor_storages) {
-        if (!found_family) {
+        if (!(is_xl || is_flux)) {
             if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
                 is_flux = true;
                 if (input_block_checked) {
@@ -1656,6 +1731,15 @@ SDVersion ModelLoader::get_sd_version() {
             if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
                 return VERSION_SD3;
             }
+            if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
+                is_wan = true;
+            }
+            if (tensor_storage.name.find("model.diffusion_model.patch_embedding.weight") != std::string::npos) {
+                patch_embedding_channels = tensor_storage.ne[3];
+            }
+            if (tensor_storage.name.find("model.diffusion_model.img_emb") != std::string::npos) {
+                has_img_emb = true;
+            }
             if (tensor_storage.name.find("model.diffusion_model.input_blocks.") != std::string::npos || tensor_storage.name.find("unet.down_blocks.") != std::string::npos) {
                 is_unet = true;
                 if (has_multiple_encoders) {
@@ -1690,11 +1774,21 @@ SDVersion ModelLoader::get_sd_version() {
         if (tensor_storage.name == "model.diffusion_model.input_blocks.0.0.weight" || tensor_storage.name == "model.diffusion_model.img_in.weight" || tensor_storage.name == "unet.conv_in.weight") {
             input_block_weight  = tensor_storage;
             input_block_checked = true;
-            if (found_family) {
+            if (is_xl || is_flux) {
                 break;
             }
         }
     }
+    if (is_wan) {
+        LOG_DEBUG("patch_embedding_channels %d", patch_embedding_channels);
+        if (patch_embedding_channels == 184320 && !has_img_emb) {
+            return VERSION_WAN2_2_I2V;
+        }
+        if (patch_embedding_channels == 147456 && !has_img_emb) {
+            return VERSION_WAN2_2_TI2V;
+        }
+        return VERSION_WAN2;
+    }
     bool is_inpaint = input_block_weight.ne[2] == 9;
     bool is_ip2p    = input_block_weight.ne[2] == 8;
     if (is_xl) {
@@ -1850,232 +1944,349 @@ std::string ModelLoader::load_t5_tokenizer_json() {
     return json_str;
 }
 
-std::vector<TensorStorage> remove_duplicates(const std::vector<TensorStorage>& vec) {
-    std::vector<TensorStorage> res;
-    std::unordered_map<std::string, size_t> name_to_index_map;
-
-    for (size_t i = 0; i < vec.size(); ++i) {
-        const std::string& current_name = vec[i].name;
-        auto it                         = name_to_index_map.find(current_name);
-
-        if (it != name_to_index_map.end()) {
-            res[it->second] = vec[i];
-        } else {
-            name_to_index_map[current_name] = i;
-            res.push_back(vec[i]);
-        }
-    }
-
-    // vec.resize(name_to_index_map.size());
-
-    return res;
+std::string ModelLoader::load_umt5_tokenizer_json() {
+    std::string json_str(reinterpret_cast<const char*>(umt5_tokenizer_json_str), sizeof(umt5_tokenizer_json_str));
+    return json_str;
 }
 
-bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend) {
-    std::vector<TensorStorage> processed_tensor_storages;
-    for (auto& tensor_storage : tensor_storages) {
-        // LOG_DEBUG("%s", name.c_str());
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p) {
+    int64_t process_time_ms = 0;
+    std::atomic<int64_t> read_time_ms(0);
+    std::atomic<int64_t> memcpy_time_ms(0);
+    std::atomic<int64_t> copy_to_backend_time_ms(0);
+    std::atomic<int64_t> convert_time_ms(0);
 
-        if (is_unused_tensor(tensor_storage.name)) {
-            continue;
+    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : (int)std::thread::hardware_concurrency();
+
+    int64_t start_time = ggml_time_ms();
+    std::vector<TensorStorage> processed_tensor_storages;
+
+    {
+        struct IndexedStorage {
+            size_t index;
+            TensorStorage ts;
+        };
+
+        std::mutex vec_mutex;
+        std::vector<IndexedStorage> all_results;
+
+        int n_threads = std::min(num_threads_to_use, (int)tensor_storages.size());
+        if (n_threads < 1) {
+            n_threads = 1;
+        }
+        std::vector<std::thread> workers;
+
+        for (int i = 0; i < n_threads; ++i) {
+            workers.emplace_back([&, thread_id = i]() {
+                std::vector<IndexedStorage> local_results;
+                std::vector<TensorStorage> temp_storages;
+
+                for (size_t j = thread_id; j < tensor_storages.size(); j += n_threads) {
+                    const auto& tensor_storage = tensor_storages[j];
+                    if (is_unused_tensor(tensor_storage.name)) {
+                        continue;
+                    }
+
+                    temp_storages.clear();
+                    preprocess_tensor(tensor_storage, temp_storages);
+
+                    for (const auto& ts : temp_storages) {
+                        local_results.push_back({j, ts});
+                    }
+                }
+
+                if (!local_results.empty()) {
+                    std::lock_guard<std::mutex> lock(vec_mutex);
+                    all_results.insert(all_results.end(),
+                                       local_results.begin(), local_results.end());
+                }
+            });
+        }
+        for (auto& w : workers) {
+            w.join();
         }
 
-        preprocess_tensor(tensor_storage, processed_tensor_storages);
+        std::unordered_map<std::string, IndexedStorage> latest_map;
+        for (auto& entry : all_results) {
+            latest_map[entry.ts.name] = entry;
+        }
+
+        processed_tensor_storages.reserve(latest_map.size());
+        for (auto& [name, entry] : latest_map) {
+            processed_tensor_storages.push_back(entry.ts);
+        }
     }
-    std::vector<TensorStorage> dedup = remove_duplicates(processed_tensor_storages);
-    processed_tensor_storages        = dedup;
+
+    process_time_ms = ggml_time_ms() - start_time;
+
+    bool success                          = true;
+    size_t total_tensors_processed        = 0;
+    const size_t total_tensors_to_process = processed_tensor_storages.size();
+    const int64_t t_start                 = ggml_time_ms();
+    int last_n_threads                    = 1;
 
-    bool success = true;
     for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
         std::string file_path = file_paths_[file_index];
         LOG_DEBUG("loading tensors from %s", file_path.c_str());
 
-        std::ifstream file(file_path, std::ios::binary);
-        if (!file.is_open()) {
-            LOG_ERROR("failed to open '%s'", file_path.c_str());
-            return false;
+        std::vector<const TensorStorage*> file_tensors;
+        for (const auto& ts : processed_tensor_storages) {
+            if (ts.file_index == file_index) {
+                file_tensors.push_back(&ts);
+            }
+        }
+        if (file_tensors.empty()) {
+            continue;
         }
 
         bool is_zip = false;
-        for (auto& tensor_storage : tensor_storages) {
-            if (tensor_storage.file_index != file_index) {
-                continue;
-            }
-            if (tensor_storage.index_in_zip >= 0) {
+        for (auto const& ts : file_tensors) {
+            if (ts->index_in_zip >= 0) {
                 is_zip = true;
                 break;
             }
         }
 
-        struct zip_t* zip = NULL;
-        if (is_zip) {
-            zip = zip_open(file_path.c_str(), 0, 'r');
-            if (zip == NULL) {
-                LOG_ERROR("failed to open zip '%s'", file_path.c_str());
-                return false;
-            }
+        int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
+        if (n_threads < 1) {
+            n_threads = 1;
+        }
+        last_n_threads = n_threads;
+
+        std::atomic<size_t> tensor_idx(0);
+        std::atomic<bool> failed(false);
+        std::vector<std::thread> workers;
+
+        for (int i = 0; i < n_threads; ++i) {
+            workers.emplace_back([&, file_path, is_zip]() {
+                std::ifstream file;
+                struct zip_t* zip = NULL;
+                if (is_zip) {
+                    zip = zip_open(file_path.c_str(), 0, 'r');
+                    if (zip == NULL) {
+                        LOG_ERROR("failed to open zip '%s'", file_path.c_str());
+                        failed = true;
+                        return;
+                    }
+                } else {
+                    file.open(file_path, std::ios::binary);
+                    if (!file.is_open()) {
+                        LOG_ERROR("failed to open '%s'", file_path.c_str());
+                        failed = true;
+                        return;
+                    }
+                }
+
+                std::vector<uint8_t> read_buffer;
+                std::vector<uint8_t> convert_buffer;
+
+                while (true) {
+                    int64_t t0, t1;
+                    size_t idx = tensor_idx.fetch_add(1);
+                    if (idx >= file_tensors.size() || failed) {
+                        break;
+                    }
+
+                    const TensorStorage& tensor_storage = *file_tensors[idx];
+                    ggml_tensor* dst_tensor             = NULL;
+
+                    t0 = ggml_time_ms();
+
+                    if (!on_new_tensor_cb(tensor_storage, &dst_tensor)) {
+                        LOG_WARN("process tensor failed: '%s'", tensor_storage.name.c_str());
+                        failed = true;
+                        break;
+                    }
+
+                    if (dst_tensor == NULL) {
+                        t1 = ggml_time_ms();
+                        read_time_ms.fetch_add(t1 - t0);
+                        continue;
+                    }
+
+                    size_t nbytes_to_read = tensor_storage.nbytes_to_read();
+
+                    auto read_data = [&](char* buf, size_t n) {
+                        if (zip != NULL) {
+                            zip_entry_openbyindex(zip, tensor_storage.index_in_zip);
+                            size_t entry_size = zip_entry_size(zip);
+                            if (entry_size != n) {
+                                int64_t t_memcpy_start;
+                                read_buffer.resize(entry_size);
+                                zip_entry_noallocread(zip, (void*)read_buffer.data(), entry_size);
+                                t_memcpy_start = ggml_time_ms();
+                                memcpy((void*)buf, (void*)(read_buffer.data() + tensor_storage.offset), n);
+                                memcpy_time_ms.fetch_add(ggml_time_ms() - t_memcpy_start);
+                            } else {
+                                zip_entry_noallocread(zip, (void*)buf, n);
+                            }
+                            zip_entry_close(zip);
+                        } else {
+                            file.seekg(tensor_storage.offset);
+                            file.read(buf, n);
+                            if (!file) {
+                                LOG_ERROR("read tensor data failed: '%s'", file_path.c_str());
+                                failed = true;
+                            }
+                        }
+                    };
+
+                    if (dst_tensor->buffer == NULL || ggml_backend_buffer_is_host(dst_tensor->buffer)) {
+                        if (tensor_storage.type == dst_tensor->type) {
+                            GGML_ASSERT(ggml_nbytes(dst_tensor) == tensor_storage.nbytes());
+                            if (tensor_storage.is_f64 || tensor_storage.is_i64) {
+                                read_buffer.resize(tensor_storage.nbytes_to_read());
+                                read_data((char*)read_buffer.data(), nbytes_to_read);
+                            } else {
+                                read_data((char*)dst_tensor->data, nbytes_to_read);
+                            }
+                            t1 = ggml_time_ms();
+                            read_time_ms.fetch_add(t1 - t0);
+
+                            t0 = ggml_time_ms();
+                            if (tensor_storage.is_bf16) {
+                                // inplace op
+                                bf16_to_f32_vec((uint16_t*)dst_tensor->data, (float*)dst_tensor->data, tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e4m3) {
+                                // inplace op
+                                f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e5m2) {
+                                // inplace op
+                                f8_e5m2_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
+                            } else if (tensor_storage.is_f64) {
+                                f64_to_f32_vec((double*)read_buffer.data(), (float*)dst_tensor->data, tensor_storage.nelements());
+                            } else if (tensor_storage.is_i64) {
+                                i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)dst_tensor->data, tensor_storage.nelements());
+                            }
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                        } else {
+                            read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
+                            read_data((char*)read_buffer.data(), nbytes_to_read);
+                            t1 = ggml_time_ms();
+                            read_time_ms.fetch_add(t1 - t0);
+
+                            t0 = ggml_time_ms();
+                            if (tensor_storage.is_bf16) {
+                                // inplace op
+                                bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e4m3) {
+                                // inplace op
+                                f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e5m2) {
+                                // inplace op
+                                f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                            } else if (tensor_storage.is_f64) {
+                                // inplace op
+                                f64_to_f32_vec((double*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                            } else if (tensor_storage.is_i64) {
+                                // inplace op
+                                i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
+                            }
+                            convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data, dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                        }
+                    } else {
+                        read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
+                        read_data((char*)read_buffer.data(), nbytes_to_read);
+                        t1 = ggml_time_ms();
+                        read_time_ms.fetch_add(t1 - t0);
+
+                        t0 = ggml_time_ms();
+                        if (tensor_storage.is_bf16) {
+                            // inplace op
+                            bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                        } else if (tensor_storage.is_f8_e4m3) {
+                            // inplace op
+                            f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                        } else if (tensor_storage.is_f8_e5m2) {
+                            // inplace op
+                            f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                        } else if (tensor_storage.is_f64) {
+                            // inplace op
+                            f64_to_f32_vec((double*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                        } else if (tensor_storage.is_i64) {
+                            // inplace op
+                            i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
+                        }
+
+                        if (tensor_storage.type == dst_tensor->type) {
+                            // copy to device memory
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                            t0 = ggml_time_ms();
+                            ggml_backend_tensor_set(dst_tensor, read_buffer.data(), 0, ggml_nbytes(dst_tensor));
+                            t1 = ggml_time_ms();
+                            copy_to_backend_time_ms.fetch_add(t1 - t0);
+                        } else {
+                            // convert first, then copy to device memory
+
+                            convert_buffer.resize(ggml_nbytes(dst_tensor));
+                            convert_tensor((void*)read_buffer.data(), tensor_storage.type, (void*)convert_buffer.data(), dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                            t0 = ggml_time_ms();
+                            ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
+                            t1 = ggml_time_ms();
+                            copy_to_backend_time_ms.fetch_add(t1 - t0);
+                        }
+                    }
+                }
+                if (zip != NULL) {
+                    zip_close(zip);
+                }
+            });
         }
 
-        std::vector<uint8_t> read_buffer;
-        std::vector<uint8_t> convert_buffer;
-
-        auto read_data = [&](const TensorStorage& tensor_storage, char* buf, size_t n) {
-            if (zip != NULL) {
-                zip_entry_openbyindex(zip, tensor_storage.index_in_zip);
-                size_t entry_size = zip_entry_size(zip);
-                if (entry_size != n) {
-                    read_buffer.resize(entry_size);
-                    zip_entry_noallocread(zip, (void*)read_buffer.data(), entry_size);
-                    memcpy((void*)buf, (void*)(read_buffer.data() + tensor_storage.offset), n);
-                } else {
-                    zip_entry_noallocread(zip, (void*)buf, n);
-                }
-                zip_entry_close(zip);
-            } else {
-                file.seekg(tensor_storage.offset);
-                file.read(buf, n);
-                if (!file) {
-                    LOG_ERROR("read tensor data failed: '%s'", file_path.c_str());
-                    return false;
-                }
-            }
-            return true;
-        };
-        int tensor_count = 0;
-        int64_t t1       = ggml_time_ms();
-        bool partial     = false;
-        for (auto& tensor_storage : processed_tensor_storages) {
-            if (tensor_storage.file_index != file_index) {
-                ++tensor_count;
-                continue;
-            }
-            ggml_tensor* dst_tensor = NULL;
-
-            success = on_new_tensor_cb(tensor_storage, &dst_tensor);
-            if (!success) {
-                LOG_WARN("process tensor failed: '%s'", tensor_storage.name.c_str());
+        while (true) {
+            size_t current_idx = tensor_idx.load();
+            if (current_idx >= file_tensors.size() || failed) {
                 break;
             }
-
-            if (dst_tensor == NULL) {
-                ++tensor_count;
-                continue;
-            }
-
-            size_t nbytes_to_read = tensor_storage.nbytes_to_read();
-
-            if (dst_tensor->buffer == NULL || ggml_backend_buffer_is_host(dst_tensor->buffer)) {
-                // for the CPU and Metal backend, we can copy directly into the tensor
-                if (tensor_storage.type == dst_tensor->type) {
-                    GGML_ASSERT(ggml_nbytes(dst_tensor) == tensor_storage.nbytes());
-                    if (tensor_storage.is_f64 || tensor_storage.is_i64) {
-                        read_buffer.resize(tensor_storage.nbytes_to_read());
-                        read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
-                    } else {
-                        read_data(tensor_storage, (char*)dst_tensor->data, nbytes_to_read);
-                    }
-
-                    if (tensor_storage.is_bf16) {
-                        // inplace op
-                        bf16_to_f32_vec((uint16_t*)dst_tensor->data, (float*)dst_tensor->data, tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e4m3) {
-                        // inplace op
-                        f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e5m2) {
-                        // inplace op
-                        f8_e5m2_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
-                    } else if (tensor_storage.is_f64) {
-                        f64_to_f32_vec((double*)read_buffer.data(), (float*)dst_tensor->data, tensor_storage.nelements());
-                    } else if (tensor_storage.is_i64) {
-                        i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)dst_tensor->data, tensor_storage.nelements());
-                    }
-                } else {
-                    read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
-                    read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
-
-                    if (tensor_storage.is_bf16) {
-                        // inplace op
-                        bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e4m3) {
-                        // inplace op
-                        f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e5m2) {
-                        // inplace op
-                        f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                    } else if (tensor_storage.is_f64) {
-                        // inplace op
-                        f64_to_f32_vec((double*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
-                    } else if (tensor_storage.is_i64) {
-                        // inplace op
-                        i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
-                    }
-
-                    convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
-                                   dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
-                }
-            } else {
-                read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
-                read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
-
-                if (tensor_storage.is_bf16) {
-                    // inplace op
-                    bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
-                } else if (tensor_storage.is_f8_e4m3) {
-                    // inplace op
-                    f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                } else if (tensor_storage.is_f8_e5m2) {
-                    // inplace op
-                    f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                } else if (tensor_storage.is_f64) {
-                    // inplace op
-                    f64_to_f32_vec((double*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
-                } else if (tensor_storage.is_i64) {
-                    // inplace op
-                    i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
-                }
-
-                if (tensor_storage.type == dst_tensor->type) {
-                    // copy to device memory
-                    ggml_backend_tensor_set(dst_tensor, read_buffer.data(), 0, ggml_nbytes(dst_tensor));
-                } else {
-                    // convert first, then copy to device memory
-                    convert_buffer.resize(ggml_nbytes(dst_tensor));
-                    convert_tensor((void*)read_buffer.data(), tensor_storage.type,
-                                   (void*)convert_buffer.data(), dst_tensor->type,
-                                   (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
-                    ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
-                }
-            }
-            size_t tensor_max = processed_tensor_storages.size();
-            int64_t t2        = ggml_time_ms();
-            pretty_progress(++tensor_count, tensor_max, (t2 - t1) / 1000.0f);
-            t1      = t2;
-            partial = tensor_count != tensor_max;
+            size_t curr_num = total_tensors_processed + current_idx;
+            pretty_progress(curr_num, total_tensors_to_process, (ggml_time_ms() - t_start) / 1000.0f / (curr_num + 1e-6f));
+            std::this_thread::sleep_for(std::chrono::milliseconds(200));
         }
 
-        if (zip != NULL) {
-            zip_close(zip);
+        for (auto& w : workers) {
+            w.join();
         }
 
-        if (partial) {
-            printf("\n");
-        }
-
-        if (!success) {
+        if (failed) {
+            success = false;
             break;
         }
+        total_tensors_processed += file_tensors.size();
+        pretty_progress(total_tensors_processed, total_tensors_to_process, (ggml_time_ms() - t_start) / 1000.0f / (total_tensors_processed + 1e-6f));
+        if (total_tensors_processed < total_tensors_to_process) {
+            printf("\n");
+        }
     }
+
+    int64_t end_time = ggml_time_ms();
+    LOG_INFO("loading tensors completed, taking %.2fs (process: %.2fs, read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
+             (end_time - start_time) / 1000.f,
+             process_time_ms / 1000.f,
+             (read_time_ms.load() / (float)last_n_threads) / 1000.f,
+             (memcpy_time_ms.load() / (float)last_n_threads) / 1000.f,
+             (convert_time_ms.load() / (float)last_n_threads) / 1000.f,
+             (copy_to_backend_time_ms.load() / (float)last_n_threads) / 1000.f);
     return success;
 }
 
 bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
-                               ggml_backend_t backend,
-                               std::set<std::string> ignore_tensors) {
+                               std::set<std::string> ignore_tensors,
+                               int n_threads) {
     std::set<std::string> tensor_names_in_file;
+    std::mutex tensor_names_mutex;
     auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
         const std::string& name = tensor_storage.name;
         // LOG_DEBUG("%s", tensor_storage.to_string().c_str());
-        tensor_names_in_file.insert(name);
+        {
+            std::lock_guard<std::mutex> lock(tensor_names_mutex);
+            tensor_names_in_file.insert(name);
+        }
 
         struct ggml_tensor* real;
         if (tensors.find(name) != tensors.end()) {
@@ -2109,7 +2320,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
         return true;
     };
 
-    bool success = load_tensors(on_new_tensor_cb, backend);
+    bool success = load_tensors(on_new_tensor_cb, n_threads);
     if (!success) {
         LOG_ERROR("load tensors from file failed");
         return false;
@@ -2140,7 +2351,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
 
 std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std::string& tensor_type_rules) {
     std::vector<std::pair<std::string, ggml_type>> result;
-    for (const auto& item : splitString(tensor_type_rules, ',')) {
+    for (const auto& item : split_string(tensor_type_rules, ',')) {
         if (item.size() == 0)
             continue;
         std::string::size_type pos = item.find('=');
@@ -2156,7 +2367,7 @@ std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std
         if (type_name == "f32") {
             tensor_type = GGML_TYPE_F32;
         } else {
-            for (size_t i = 0; i < SD_TYPE_COUNT; i++) {
+            for (size_t i = 0; i < GGML_TYPE_COUNT; i++) {
                 auto trait = ggml_get_type_traits((ggml_type)i);
                 if (trait->to_float && trait->type_size && type_name == trait->type_name) {
                     tensor_type = (ggml_type)i;
@@ -2253,7 +2464,7 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
         return true;
     };
 
-    bool success = load_tensors(on_new_tensor_cb, backend);
+    bool success = load_tensors(on_new_tensor_cb);
     ggml_backend_free(backend);
     LOG_INFO("load tensors done");
     LOG_INFO("trying to save tensors to %s", file_path.c_str());

model.h

@@ -31,23 +31,12 @@ enum SDVersion {
     VERSION_SD3,
     VERSION_FLUX,
     VERSION_FLUX_FILL,
+    VERSION_WAN2,
+    VERSION_WAN2_2_I2V,
+    VERSION_WAN2_2_TI2V,
     VERSION_COUNT,
 };
 
-static inline bool sd_version_is_flux(SDVersion version) {
-    if (version == VERSION_FLUX || version == VERSION_FLUX_FILL) {
-        return true;
-    }
-    return false;
-}
-
-static inline bool sd_version_is_sd3(SDVersion version) {
-    if (version == VERSION_SD3) {
-        return true;
-    }
-    return false;
-}
-
 static inline bool sd_version_is_sd1(SDVersion version) {
     if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT || version == VERSION_SD1_PIX2PIX) {
         return true;
@@ -69,6 +58,27 @@ static inline bool sd_version_is_sdxl(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_sd3(SDVersion version) {
+    if (version == VERSION_SD3) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_flux(SDVersion version) {
+    if (version == VERSION_FLUX || version == VERSION_FLUX_FILL) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_wan(SDVersion version) {
+    if (version == VERSION_WAN2 || version == VERSION_WAN2_2_I2V || version == VERSION_WAN2_2_TI2V) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_inpaint(SDVersion version) {
     if (version == VERSION_SD1_INPAINT || version == VERSION_SD2_INPAINT || version == VERSION_SDXL_INPAINT || version == VERSION_FLUX_FILL) {
         return true;
@@ -77,7 +87,7 @@ static inline bool sd_version_is_inpaint(SDVersion version) {
 }
 
 static inline bool sd_version_is_dit(SDVersion version) {
-    if (sd_version_is_flux(version) || sd_version_is_sd3(version)) {
+    if (sd_version_is_flux(version) || sd_version_is_sd3(version) || sd_version_is_wan(version)) {
         return true;
     }
     return false;
@@ -109,11 +119,11 @@ struct TensorStorage {
 
     size_t file_index = 0;
     int index_in_zip  = -1;  // >= means stored in a zip file
-    size_t offset     = 0;   // offset in file
+    uint64_t offset   = 0;   // offset in file
 
     TensorStorage() = default;
 
-    TensorStorage(const std::string& name, ggml_type type, int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
+    TensorStorage(const std::string& name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
         : name(name), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
         for (int i = 0; i < n_dims; i++) {
             this->ne[i] = ne[i];
@@ -154,10 +164,10 @@ struct TensorStorage {
 
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
-        size_t chunk_size = nbytes_to_read() / n;
+        uint64_t chunk_size = nbytes_to_read() / n;
         // printf("%d/%d\n", chunk_size, nbytes_to_read());
         reverse_ne();
-        for (int i = 0; i < n; i++) {
+        for (size_t i = 0; i < n; i++) {
             TensorStorage chunk_i = *this;
             chunk_i.ne[0]         = ne[0] / n;
             chunk_i.offset        = offset + i * chunk_size;
@@ -207,6 +217,8 @@ struct TensorStorage {
 
 typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
 
+typedef std::map<std::string, enum ggml_type> String2GGMLType;
+
 class ModelLoader {
 protected:
     std::vector<std::string> file_paths_;
@@ -225,7 +237,7 @@ protected:
     bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
 
 public:
-    std::map<std::string, enum ggml_type> tensor_storages_types;
+    String2GGMLType tensor_storages_types;
 
     bool init_from_file(const std::string& file_path, const std::string& prefix = "");
     bool model_is_unet();
@@ -235,10 +247,10 @@ public:
     ggml_type get_diffusion_model_wtype();
     ggml_type get_vae_wtype();
     void set_wtype_override(ggml_type wtype, std::string prefix = "");
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend);
+    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0);
     bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
-                      ggml_backend_t backend,
-                      std::set<std::string> ignore_tensors = {});
+                      std::set<std::string> ignore_tensors = {},
+                      int n_threads                        = 0);
 
     bool save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules);
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
@@ -247,6 +259,7 @@ public:
 
     static std::string load_merges();
     static std::string load_t5_tokenizer_json();
+    static std::string load_umt5_tokenizer_json();
 };
 
 #endif  // __MODEL_H__

pmid.hpp

@@ -42,41 +42,6 @@ public:
     }
 };
 
-/*
-class QFormerPerceiver(nn.Module):
-    def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
-        super().__init__()
-
-        self.num_tokens = num_tokens
-        self.cross_attention_dim = cross_attention_dim
-        self.use_residual = use_residual
-        print(cross_attention_dim*num_tokens)
-        self.token_proj = nn.Sequential(
-            nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
-            nn.GELU(),
-            nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
-        )
-        self.token_norm = nn.LayerNorm(cross_attention_dim)
-        self.perceiver_resampler = FacePerceiverResampler(
-            dim=cross_attention_dim,
-            depth=4,
-            dim_head=128,
-            heads=cross_attention_dim // 128,
-            embedding_dim=embedding_dim,
-            output_dim=cross_attention_dim,
-            ff_mult=4,
-        )
-
-    def forward(self, x, last_hidden_state):
-        x = self.token_proj(x)
-        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
-        x = self.token_norm(x) # cls token
-        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
-        if self.use_residual: # TODO: if use_residual is not true
-            out = x + 1.0 * out
-        return out
-*/
-
 struct PMFeedForward : public GGMLBlock {
     // network hparams
     int dim;
@@ -122,17 +87,8 @@ public:
         int64_t ne[4];
         for (int i = 0; i < 4; ++i)
             ne[i] = x->ne[i];
-        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 0: ");
-        // printf("heads = %d \n", heads);
-        // x = ggml_view_4d(ctx, x, x->ne[0], x->ne[1], heads, x->ne[2]/heads,
-        //                          x->nb[1], x->nb[2], x->nb[3], 0);
         x = ggml_reshape_4d(ctx, x, x->ne[0] / heads, heads, x->ne[1], x->ne[2]);
-        // x = ggml_view_4d(ctx, x, x->ne[0]/heads, heads, x->ne[1], x->ne[2],
-        //                          x->nb[1], x->nb[2], x->nb[3], 0);
-        // x = ggml_cont(ctx, x);
         x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
-        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 1: ");
-        // x  = ggml_reshape_4d(ctx, x, ne[0], heads, ne[1], ne[2]/heads);
         return x;
     }
 
@@ -269,17 +225,6 @@ public:
             4));
     }
 
-    /*
-    def forward(self, x, last_hidden_state):
-        x = self.token_proj(x)
-        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
-        x = self.token_norm(x) # cls token
-        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
-        if self.use_residual: # TODO: if use_residual is not true
-            out = x + 1.0 * out
-        return out
-    */
-
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* last_hidden_state) {
@@ -299,113 +244,6 @@ public:
     }
 };
 
-/*
-class FacePerceiverResampler(torch.nn.Module):
-    def __init__(
-        self,
-        *,
-        dim=768,
-        depth=4,
-        dim_head=64,
-        heads=16,
-        embedding_dim=1280,
-        output_dim=768,
-        ff_mult=4,
-    ):
-        super().__init__()
-
-        self.proj_in = torch.nn.Linear(embedding_dim, dim)
-        self.proj_out = torch.nn.Linear(dim, output_dim)
-        self.norm_out = torch.nn.LayerNorm(output_dim)
-        self.layers = torch.nn.ModuleList([])
-        for _ in range(depth):
-            self.layers.append(
-                torch.nn.ModuleList(
-                    [
-                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
-                        FeedForward(dim=dim, mult=ff_mult),
-                    ]
-                )
-            )
-
-    def forward(self, latents, x):
-        x = self.proj_in(x)
-        for attn, ff in self.layers:
-            latents = attn(x, latents) + latents
-            latents = ff(latents) + latents
-        latents = self.proj_out(latents)
-        return self.norm_out(latents)
-*/
-
-/*
-
-def FeedForward(dim, mult=4):
-    inner_dim = int(dim * mult)
-    return nn.Sequential(
-        nn.LayerNorm(dim),
-        nn.Linear(dim, inner_dim, bias=False),
-        nn.GELU(),
-        nn.Linear(inner_dim, dim, bias=False),
-    )
-
-def reshape_tensor(x, heads):
-    bs, length, width = x.shape
-    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
-    x = x.view(bs, length, heads, -1)
-    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
-    x = x.transpose(1, 2)
-    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
-    x = x.reshape(bs, heads, length, -1)
-    return x
-
-class PerceiverAttention(nn.Module):
-    def __init__(self, *, dim, dim_head=64, heads=8):
-        super().__init__()
-        self.scale = dim_head**-0.5
-        self.dim_head = dim_head
-        self.heads = heads
-        inner_dim = dim_head * heads
-
-        self.norm1 = nn.LayerNorm(dim)
-        self.norm2 = nn.LayerNorm(dim)
-
-        self.to_q = nn.Linear(dim, inner_dim, bias=False)
-        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
-        self.to_out = nn.Linear(inner_dim, dim, bias=False)
-
-    def forward(self, x, latents):
-        """
-        Args:
-            x (torch.Tensor): image features
-                shape (b, n1, D)
-            latent (torch.Tensor): latent features
-                shape (b, n2, D)
-        """
-        x = self.norm1(x)
-        latents = self.norm2(latents)
-
-        b, l, _ = latents.shape
-
-        q = self.to_q(latents)
-        kv_input = torch.cat((x, latents), dim=-2)
-        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
-
-        q = reshape_tensor(q, self.heads)
-        k = reshape_tensor(k, self.heads)
-        v = reshape_tensor(v, self.heads)
-
-        # attention
-        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
-        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
-        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
-        out = weight @ v
-
-        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
-
-        return self.to_out(out)
-
-*/
-
 struct FuseModule : public GGMLBlock {
     // network hparams
     int embed_dim;
@@ -425,31 +263,13 @@ public:
         auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
         auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
 
-        // print_ggml_tensor(id_embeds, true, "Fuseblock id_embeds: ");
-        // print_ggml_tensor(prompt_embeds, true, "Fuseblock prompt_embeds: ");
-
-        // auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
-        // auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
-        // print_ggml_tensor(id_embeds0, true, "Fuseblock id_embeds0: ");
-        // print_ggml_tensor(prompt_embeds0, true, "Fuseblock prompt_embeds0: ");
-        // concat is along dim 2
-        // auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
         auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds, id_embeds, 0);
-        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 0: ");
-        // stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
-        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
-        // stacked_id_embeds = mlp1.forward(ctx, stacked_id_embeds);
-        // stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
-        // stacked_id_embeds = mlp2.forward(ctx, stacked_id_embeds);
-        // stacked_id_embeds = ggml_nn_layer_norm(ctx, stacked_id_embeds, ln_w, ln_b);
 
         stacked_id_embeds = mlp1->forward(ctx, stacked_id_embeds);
         stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
         stacked_id_embeds = mlp2->forward(ctx, stacked_id_embeds);
         stacked_id_embeds = layer_norm->forward(ctx, stacked_id_embeds);
 
-        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
-
         return stacked_id_embeds;
     }
 
@@ -464,21 +284,14 @@ public:
 
         struct ggml_tensor* valid_id_embeds = id_embeds;
         // # slice out the image token embeddings
-        // print_ggml_tensor(class_tokens_mask_pos, false);
         ggml_set_name(class_tokens_mask_pos, "class_tokens_mask_pos");
         ggml_set_name(prompt_embeds, "prompt_embeds");
-        // print_ggml_tensor(valid_id_embeds, true, "valid_id_embeds");
-        // print_ggml_tensor(class_tokens_mask_pos, true, "class_tokens_mask_pos");
         struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx, prompt_embeds, class_tokens_mask_pos);
         ggml_set_name(image_token_embeds, "image_token_embeds");
         valid_id_embeds                       = ggml_reshape_2d(ctx, valid_id_embeds, valid_id_embeds->ne[0],
                                                                 ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
         struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
 
-        // stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
-        // print_ggml_tensor(stacked_id_embeds, true, "AA stacked_id_embeds");
-        // print_ggml_tensor(left, true, "AA left");
-        // print_ggml_tensor(right, true, "AA right");
         if (left && right) {
             stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
             stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
@@ -487,15 +300,12 @@ public:
         } else if (right) {
             stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
         }
-        // print_ggml_tensor(stacked_id_embeds, true, "BB stacked_id_embeds");
-        // stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
-        // print_ggml_tensor(stacked_id_embeds, true, "CC stacked_id_embeds");
+
         class_tokens_mask                         = ggml_cont(ctx, ggml_transpose(ctx, class_tokens_mask));
         class_tokens_mask                         = ggml_repeat(ctx, class_tokens_mask, prompt_embeds);
         prompt_embeds                             = ggml_mul(ctx, prompt_embeds, class_tokens_mask);
         struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx, prompt_embeds, stacked_id_embeds);
         ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
-        // print_ggml_tensor(updated_prompt_embeds, true, "updated_prompt_embeds: ");
         return updated_prompt_embeds;
     }
 };
@@ -508,6 +318,7 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* id_pixel_values,
                                 struct ggml_tensor* prompt_embeds,
                                 struct ggml_tensor* class_tokens_mask,
@@ -520,9 +331,9 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
         auto visual_projection_2 = std::dynamic_pointer_cast<Linear>(blocks["visual_projection_2"]);
         auto fuse_module         = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
 
-        struct ggml_tensor* shared_id_embeds = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
-        struct ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);    // [N, proj_dim(768)]
-        struct ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);  // [N, 1280]
+        struct ggml_tensor* shared_id_embeds = vision_model->forward(ctx, backend, id_pixel_values);  // [N, hidden_size]
+        struct ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);     // [N, proj_dim(768)]
+        struct ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);   // [N, 1280]
 
         id_embeds   = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
         id_embeds_2 = ggml_cont(ctx, ggml_permute(ctx, id_embeds_2, 2, 0, 1, 3));
@@ -550,35 +361,13 @@ struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionMo
           num_tokens(2) {
         blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
         blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
-        /*
-        cross_attention_dim = 2048
-        # projection
-        self.num_tokens = 2
-        self.cross_attention_dim = cross_attention_dim
-        self.qformer_perceiver = QFormerPerceiver(
-                                    id_embeddings_dim,
-                                    cross_attention_dim,
-                                    self.num_tokens,
-                                )*/
-        blocks["qformer_perceiver"] = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
-                                                                                      cross_attention_dim,
-                                                                                      num_tokens));
+        blocks["qformer_perceiver"]   = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
+                                                                                        cross_attention_dim,
+                                                                                        num_tokens));
     }
 
-    /*
-    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
-        b, num_inputs, c, h, w = id_pixel_values.shape
-        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
-
-        last_hidden_state = self.vision_model(id_pixel_values)[0]
-        id_embeds = id_embeds.view(b * num_inputs, -1)
-
-        id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
-        id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
-        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
-    */
-
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* id_pixel_values,
                                 struct ggml_tensor* prompt_embeds,
                                 struct ggml_tensor* class_tokens_mask,
@@ -592,7 +381,7 @@ struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionMo
         auto qformer_perceiver = std::dynamic_pointer_cast<QFormerPerceiver>(blocks["qformer_perceiver"]);
 
         // struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
-        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
+        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, backend, id_pixel_values, false);  // [N, hidden_size]
         id_embeds                             = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
 
         struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
@@ -623,8 +412,14 @@ public:
     std::vector<float> zeros_right;
 
 public:
-    PhotoMakerIDEncoder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix, SDVersion version = VERSION_SDXL, PMVersion pm_v = PM_VERSION_1, float sty = 20.f)
-        : GGMLRunner(backend),
+    PhotoMakerIDEncoder(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types,
+                        const std::string prefix,
+                        SDVersion version = VERSION_SDXL,
+                        PMVersion pm_v    = PM_VERSION_1,
+                        float sty         = 20.f)
+        : GGMLRunner(backend, offload_params_to_cpu),
           version(version),
           pm_version(pm_v),
           style_strength(sty) {
@@ -736,6 +531,7 @@ public:
         struct ggml_tensor* updated_prompt_embeds = NULL;
         if (pm_version == PM_VERSION_1)
             updated_prompt_embeds = id_encoder.forward(ctx0,
+                                                       runtime_backend,
                                                        id_pixel_values_d,
                                                        prompt_embeds_d,
                                                        class_tokens_mask_d,
@@ -743,6 +539,7 @@ public:
                                                        left, right);
         else if (pm_version == PM_VERSION_2)
             updated_prompt_embeds = id_encoder2.forward(ctx0,
+                                                        runtime_backend,
                                                         id_pixel_values_d,
                                                         prompt_embeds_d,
                                                         class_tokens_mask_d,
@@ -780,10 +577,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
     bool applied     = false;
 
     PhotoMakerIDEmbed(ggml_backend_t backend,
+                      bool offload_params_to_cpu,
                       ModelLoader* ml,
                       const std::string& file_path = "",
                       const std::string& prefix    = "")
-        : file_path(file_path), GGMLRunner(backend), model_loader(ml) {
+        : file_path(file_path), GGMLRunner(backend, offload_params_to_cpu), model_loader(ml) {
         if (!model_loader->init_from_file(file_path, prefix)) {
             load_failed = true;
         }
@@ -823,11 +621,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
             return true;
         };
 
-        model_loader->load_tensors(on_new_tensor_cb, backend);
+        model_loader->load_tensors(on_new_tensor_cb);
         alloc_params_buffer();
 
         dry_run = false;
-        model_loader->load_tensors(on_new_tensor_cb, backend);
+        model_loader->load_tensors(on_new_tensor_cb);
 
         LOG_DEBUG("finished loading PhotoMaker ID Embeds ");
         return true;

preprocessing.hpp

@@ -162,16 +162,16 @@ void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float lo
     }
 }
 
-uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
+bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
     struct ggml_init_params params;
-    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10
+    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10MB
     params.mem_buffer             = NULL;
     params.no_alloc               = false;
     struct ggml_context* work_ctx = ggml_init(params);
 
     if (!work_ctx) {
         LOG_ERROR("ggml_init() failed");
-        return NULL;
+        return false;
     }
 
     float kX[9] = {
@@ -192,8 +192,8 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_thresh
     struct ggml_tensor* sf_ky = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
     memcpy(sf_ky->data, kY, ggml_nbytes(sf_ky));
     gaussian_kernel(gkernel);
-    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 3, 1);
-    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 1, 1);
+    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 3, 1);
+    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 1, 1);
     struct ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
     struct ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
     struct ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
@@ -209,8 +209,8 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_thresh
     non_max_supression(image_gray, G, tetha);
     threshold_hystersis(image_gray, high_threshold, low_threshold, weak, strong);
     // to RGB channels
-    for (int iy = 0; iy < height; iy++) {
-        for (int ix = 0; ix < width; ix++) {
+    for (int iy = 0; iy < img.height; iy++) {
+        for (int ix = 0; ix < img.width; ix++) {
             float gray = ggml_tensor_get_f32(image_gray, ix, iy);
             gray       = inverse ? 1.0f - gray : gray;
             ggml_tensor_set_f32(image, gray, ix, iy);
@@ -218,10 +218,11 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_thresh
             ggml_tensor_set_f32(image, gray, ix, iy, 2);
         }
     }
-    free(img);
     uint8_t* output = sd_tensor_to_image(image);
+    free(img.data);
+    img.data = output;
     ggml_free(work_ctx);
-    return output;
+    return true;
 }
 
 #endif  // __PREPROCESSING_HPP__

rope.hpp

@@ -0,0 +1,261 @@
+#ifndef __ROPE_HPP__
+#define __ROPE_HPP__
+
+#include <vector>
+#include "ggml_extend.hpp"
+
+struct Rope {
+    template <class T>
+    static std::vector<T> linspace(T start, T end, int num) {
+        std::vector<T> result(num);
+        if (num == 1) {
+            result[0] = start;
+            return result;
+        }
+        T step = (end - start) / (num - 1);
+        for (int i = 0; i < num; ++i) {
+            result[i] = start + i * step;
+        }
+        return result;
+    }
+
+    static std::vector<std::vector<float>> transpose(const std::vector<std::vector<float>>& mat) {
+        int rows = mat.size();
+        int cols = mat[0].size();
+        std::vector<std::vector<float>> transposed(cols, std::vector<float>(rows));
+        for (int i = 0; i < rows; ++i) {
+            for (int j = 0; j < cols; ++j) {
+                transposed[j][i] = mat[i][j];
+            }
+        }
+        return transposed;
+    }
+
+    static std::vector<float> flatten(const std::vector<std::vector<float>>& vec) {
+        std::vector<float> flat_vec;
+        for (const auto& sub_vec : vec) {
+            flat_vec.insert(flat_vec.end(), sub_vec.begin(), sub_vec.end());
+        }
+        return flat_vec;
+    }
+
+    static std::vector<std::vector<float>> rope(const std::vector<float>& pos, int dim, int theta) {
+        assert(dim % 2 == 0);
+        int half_dim = dim / 2;
+
+        std::vector<float> scale = linspace(0.f, (dim * 1.f - 2) / dim, half_dim);
+
+        std::vector<float> omega(half_dim);
+        for (int i = 0; i < half_dim; ++i) {
+            omega[i] = 1.0 / std::pow(theta, scale[i]);
+        }
+
+        int pos_size = pos.size();
+        std::vector<std::vector<float>> out(pos_size, std::vector<float>(half_dim));
+        for (int i = 0; i < pos_size; ++i) {
+            for (int j = 0; j < half_dim; ++j) {
+                out[i][j] = pos[i] * omega[j];
+            }
+        }
+
+        std::vector<std::vector<float>> result(pos_size, std::vector<float>(half_dim * 4));
+        for (int i = 0; i < pos_size; ++i) {
+            for (int j = 0; j < half_dim; ++j) {
+                result[i][4 * j]     = std::cos(out[i][j]);
+                result[i][4 * j + 1] = -std::sin(out[i][j]);
+                result[i][4 * j + 2] = std::sin(out[i][j]);
+                result[i][4 * j + 3] = std::cos(out[i][j]);
+            }
+        }
+
+        return result;
+    }
+
+    // Generate IDs for image patches and text
+    static std::vector<std::vector<float>> gen_txt_ids(int bs, int context_len) {
+        return std::vector<std::vector<float>>(bs * context_len, std::vector<float>(3, 0.0));
+    }
+
+    static std::vector<std::vector<float>> gen_img_ids(int h, int w, int patch_size, int bs, int index = 0, int h_offset = 0, int w_offset = 0) {
+        int h_len = (h + (patch_size / 2)) / patch_size;
+        int w_len = (w + (patch_size / 2)) / patch_size;
+
+        std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(3, 0.0));
+
+        std::vector<float> row_ids = linspace<float>(h_offset, h_len - 1 + h_offset, h_len);
+        std::vector<float> col_ids = linspace<float>(w_offset, w_len - 1 + w_offset, w_len);
+
+        for (int i = 0; i < h_len; ++i) {
+            for (int j = 0; j < w_len; ++j) {
+                img_ids[i * w_len + j][0] = index;
+                img_ids[i * w_len + j][1] = row_ids[i];
+                img_ids[i * w_len + j][2] = col_ids[j];
+            }
+        }
+
+        std::vector<std::vector<float>> img_ids_repeated(bs * img_ids.size(), std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < img_ids.size(); ++j) {
+                img_ids_repeated[i * img_ids.size() + j] = img_ids[j];
+            }
+        }
+        return img_ids_repeated;
+    }
+
+    static std::vector<std::vector<float>> concat_ids(const std::vector<std::vector<float>>& a,
+                                                      const std::vector<std::vector<float>>& b,
+                                                      int bs) {
+        size_t a_len = a.size() / bs;
+        size_t b_len = b.size() / bs;
+        std::vector<std::vector<float>> ids(a.size() + b.size(), std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < a_len; ++j) {
+                ids[i * (a_len + b_len) + j] = a[i * a_len + j];
+            }
+            for (int j = 0; j < b_len; ++j) {
+                ids[i * (a_len + b_len) + a_len + j] = b[i * b_len + j];
+            }
+        }
+        return ids;
+    }
+
+    static std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
+                                       int bs,
+                                       int theta,
+                                       const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> trans_ids = transpose(ids);
+        size_t pos_len                            = ids.size() / bs;
+        int num_axes                              = axes_dim.size();
+        // for (int i = 0; i < pos_len; i++) {
+        //     std::cout << trans_ids[0][i] << " " << trans_ids[1][i] << " " << trans_ids[2][i] << std::endl;
+        // }
+
+        int emb_dim = 0;
+        for (int d : axes_dim)
+            emb_dim += d / 2;
+
+        std::vector<std::vector<float>> emb(bs * pos_len, std::vector<float>(emb_dim * 2 * 2, 0.0));
+        int offset = 0;
+        for (int i = 0; i < num_axes; ++i) {
+            std::vector<std::vector<float>> rope_emb = rope(trans_ids[i], axes_dim[i], theta);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
+            for (int b = 0; b < bs; ++b) {
+                for (int j = 0; j < pos_len; ++j) {
+                    for (int k = 0; k < rope_emb[0].size(); ++k) {
+                        emb[b * pos_len + j][offset + k] = rope_emb[j][k];
+                    }
+                }
+            }
+            offset += rope_emb[0].size();
+        }
+
+        return flatten(emb);
+    }
+
+    static std::vector<std::vector<float>> gen_flux_ids(int h,
+                                                        int w,
+                                                        int patch_size,
+                                                        int bs,
+                                                        int context_len,
+                                                        std::vector<ggml_tensor*> ref_latents,
+                                                        bool increase_ref_index) {
+        auto txt_ids = gen_txt_ids(bs, context_len);
+        auto img_ids = gen_img_ids(h, w, patch_size, bs);
+
+        auto ids               = concat_ids(txt_ids, img_ids, bs);
+        uint64_t curr_h_offset = 0;
+        uint64_t curr_w_offset = 0;
+        int index              = 1;
+        for (ggml_tensor* ref : ref_latents) {
+            uint64_t h_offset = 0;
+            uint64_t w_offset = 0;
+            if (!increase_ref_index) {
+                if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) {
+                    w_offset = curr_w_offset;
+                } else {
+                    h_offset = curr_h_offset;
+                }
+            }
+
+            auto ref_ids = gen_img_ids(ref->ne[1], ref->ne[0], patch_size, bs, index, h_offset, w_offset);
+            ids          = concat_ids(ids, ref_ids, bs);
+
+            if (increase_ref_index) {
+                index++;
+            }
+
+            curr_h_offset = std::max(curr_h_offset, ref->ne[1] + h_offset);
+            curr_w_offset = std::max(curr_w_offset, ref->ne[0] + w_offset);
+        }
+        return ids;
+    }
+
+    // Generate flux positional embeddings
+    static std::vector<float> gen_flux_pe(int h,
+                                          int w,
+                                          int patch_size,
+                                          int bs,
+                                          int context_len,
+                                          std::vector<ggml_tensor*> ref_latents,
+                                          bool increase_ref_index,
+                                          int theta,
+                                          const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_flux_ids(h, w, patch_size, bs, context_len, ref_latents, increase_ref_index);
+        return embed_nd(ids, bs, theta, axes_dim);
+    }
+
+    static 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>(t_offset, t_len - 1 + t_offset, t_len);
+        std::vector<float> h_ids = linspace<float>(h_offset, h_len - 1 + h_offset, h_len);
+        std::vector<float> w_ids = linspace<float>(w_offset, 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 std::vector<float> gen_wan_pe(int t,
+                                         int h,
+                                         int w,
+                                         int pt,
+                                         int ph,
+                                         int pw,
+                                         int bs,
+                                         int theta,
+                                         const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_vid_ids(t, h, w, pt, ph, pw, bs);
+        return embed_nd(ids, bs, theta, axes_dim);
+    }
+};  // struct Rope
+
+#endif  // __ROPE_HPP__

stable-diffusion.h

@@ -35,7 +35,7 @@ enum rng_type_t {
 };
 
 enum sample_method_t {
-    EULER_A,
+    SAMPLE_METHOD_DEFAULT,
     EULER,
     HEUN,
     DPM2,
@@ -47,16 +47,18 @@ enum sample_method_t {
     LCM,
     DDIM_TRAILING,
     TCD,
+    EULER_A,
     SAMPLE_METHOD_COUNT
 };
 
-enum schedule_t {
+enum scheduler_t {
     DEFAULT,
     DISCRETE,
     KARRAS,
     EXPONENTIAL,
     AYS,
     GITS,
+    SMOOTHSTEP,
     SCHEDULE_COUNT
 };
 
@@ -101,7 +103,8 @@ enum sd_type_t {
     // SD_TYPE_IQ4_NL_4_4 = 36,
     // SD_TYPE_IQ4_NL_4_8 = 37,
     // SD_TYPE_IQ4_NL_8_8 = 38,
-    SD_TYPE_COUNT = 39,
+    SD_TYPE_MXFP4 = 39,  // MXFP4 (1 block)
+    SD_TYPE_COUNT = 40,
 };
 
 enum sd_log_level_t {
@@ -111,32 +114,45 @@ enum sd_log_level_t {
     SD_LOG_ERROR
 };
 
+typedef struct {
+    bool enabled;
+    int tile_size_x;
+    int tile_size_y;
+    float target_overlap;
+    float rel_size_x;
+    float rel_size_y;
+} sd_tiling_params_t;
+
 typedef struct {
     const char* model_path;
     const char* clip_l_path;
     const char* clip_g_path;
+    const char* clip_vision_path;
     const char* t5xxl_path;
     const char* diffusion_model_path;
+    const char* high_noise_diffusion_model_path;
     const char* vae_path;
     const char* taesd_path;
     const char* control_net_path;
     const char* lora_model_dir;
     const char* embedding_dir;
-    const char* stacked_id_embed_dir;
+    const char* photo_maker_path;
     bool vae_decode_only;
-    bool vae_tiling;
     bool free_params_immediately;
     int n_threads;
     enum sd_type_t wtype;
     enum rng_type_t rng_type;
-    enum schedule_t schedule;
+    bool offload_params_to_cpu;
     bool keep_clip_on_cpu;
     bool keep_control_net_on_cpu;
     bool keep_vae_on_cpu;
     bool diffusion_flash_attn;
+    bool diffusion_conv_direct;
+    bool vae_conv_direct;
     bool chroma_use_dit_mask;
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
+    float flow_shift;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -157,48 +173,64 @@ typedef struct {
 typedef struct {
     float txt_cfg;
     float img_cfg;
-    float min_cfg;
     float distilled_guidance;
     sd_slg_params_t slg;
 } sd_guidance_params_t;
 
 typedef struct {
-    const char* prompt;
-    const char* negative_prompt;
-    int clip_skip;
     sd_guidance_params_t guidance;
-    sd_image_t init_image;
-    sd_image_t* ref_images;
-    int ref_images_count;
-    sd_image_t mask_image;
-    int width;
-    int height;
+    enum scheduler_t scheduler;
     enum sample_method_t sample_method;
     int sample_steps;
     float eta;
+} sd_sample_params_t;
+
+typedef struct {
+    sd_image_t* id_images;
+    int id_images_count;
+    const char* id_embed_path;
+    float style_strength;
+} sd_pm_params_t;  // photo maker
+
+typedef struct {
+    const char* prompt;
+    const char* negative_prompt;
+    int clip_skip;
+    sd_image_t init_image;
+    sd_image_t* ref_images;
+    int ref_images_count;
+    bool increase_ref_index;
+    sd_image_t mask_image;
+    int width;
+    int height;
+    sd_sample_params_t sample_params;
     float strength;
     int64_t seed;
     int batch_count;
-    const sd_image_t* control_cond;
+    sd_image_t control_image;
     float control_strength;
-    float style_strength;
     bool normalize_input;
-    const char* input_id_images_path;
+    sd_pm_params_t pm_params;
+    sd_tiling_params_t vae_tiling_params;
 } sd_img_gen_params_t;
 
 typedef struct {
+    const char* prompt;
+    const char* negative_prompt;
+    int clip_skip;
     sd_image_t init_image;
+    sd_image_t end_image;
+    sd_image_t* control_frames;
+    int control_frames_size;
     int width;
     int height;
-    sd_guidance_params_t guidance;
-    enum sample_method_t sample_method;
-    int sample_steps;
+    sd_sample_params_t sample_params;
+    sd_sample_params_t high_noise_sample_params;
+    float moe_boundary;
     float strength;
     int64_t seed;
     int video_frames;
-    int motion_bucket_id;
-    int fps;
-    float augmentation_level;
+    float vace_strength;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
@@ -217,29 +249,37 @@ SD_API const char* sd_rng_type_name(enum rng_type_t rng_type);
 SD_API enum rng_type_t str_to_rng_type(const char* str);
 SD_API const char* sd_sample_method_name(enum sample_method_t sample_method);
 SD_API enum sample_method_t str_to_sample_method(const char* str);
-SD_API const char* sd_schedule_name(enum schedule_t schedule);
-SD_API enum schedule_t str_to_schedule(const char* str);
+SD_API const char* sd_schedule_name(enum scheduler_t scheduler);
+SD_API enum scheduler_t str_to_schedule(const char* str);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);
 
 SD_API sd_ctx_t* new_sd_ctx(const sd_ctx_params_t* sd_ctx_params);
 SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
+SD_API enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx);
+
+SD_API void sd_sample_params_init(sd_sample_params_t* sample_params);
+SD_API char* sd_sample_params_to_str(const sd_sample_params_t* sample_params);
 
 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 void sd_vid_gen_params_init(sd_vid_gen_params_t* sd_vid_gen_params);
-SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* sd_vid_gen_params);  // broken
+SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* sd_vid_gen_params, int* num_frames_out);
 
 typedef struct upscaler_ctx_t upscaler_ctx_t;
 
 SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
+                                        bool offload_params_to_cpu,
+                                        bool direct,
                                         int n_threads);
 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 sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,
+                          sd_image_t input_image,
+                          uint32_t upscale_factor);
 
 SD_API bool convert(const char* input_path,
                     const char* vae_path,
@@ -247,14 +287,12 @@ SD_API bool convert(const char* input_path,
                     enum sd_type_t output_type,
                     const char* tensor_type_rules);
 
-SD_API uint8_t* preprocess_canny(uint8_t* img,
-                                 int width,
-                                 int height,
-                                 float high_threshold,
-                                 float low_threshold,
-                                 float weak,
-                                 float strong,
-                                 bool inverse);
+SD_API bool preprocess_canny(sd_image_t image,
+                             float high_threshold,
+                             float low_threshold,
+                             float weak,
+                             float strong,
+                             bool inverse);
 
 #ifdef __cplusplus
 }

t5.hpp

@@ -124,7 +124,10 @@ protected:
                 return;
             }
             std::string piece = item[0];
-            float score       = item[1];
+            if (piece.empty()) {
+                piece = "<empty_token>";
+            }
+            float score = item[1];
             piece_score_pairs.emplace_back(piece, score);
         }
     }
@@ -147,6 +150,7 @@ protected:
         std::vector<const char*> key(pieces->size());
         std::vector<int> value(pieces->size());
         for (size_t i = 0; i < pieces->size(); ++i) {
+            // LOG_DEBUG("%s %d", (*pieces)[i].first.c_str(), (*pieces)[i].second);
             key[i]   = (*pieces)[i].first.data();  // sorted piece.
             value[i] = (*pieces)[i].second;        // vocab_id
         }
@@ -335,9 +339,9 @@ protected:
     }
 
 public:
-    explicit T5UniGramTokenizer(const std::string& json_str = "") {
-        if (json_str.size() != 0) {
-            InitializePieces(json_str);
+    explicit T5UniGramTokenizer(bool is_umt5 = false) {
+        if (is_umt5) {
+            InitializePieces(ModelLoader::load_umt5_tokenizer_json());
         } else {
             InitializePieces(ModelLoader::load_t5_tokenizer_json());
         }
@@ -457,8 +461,8 @@ protected:
     int64_t hidden_size;
     float eps;
 
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
-        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F32;
         params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
     }
 
@@ -574,6 +578,7 @@ public:
 
     // x: [N, n_token, model_dim]
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                ggml_backend_t backend,
                                                                 struct ggml_tensor* x,
                                                                 struct ggml_tensor* past_bias                = NULL,
                                                                 struct ggml_tensor* mask                     = NULL,
@@ -604,7 +609,7 @@ public:
 
         k = ggml_scale_inplace(ctx, k, sqrt(d_head));
 
-        x = ggml_nn_attention_ext(ctx, q, k, v, num_heads, mask);  // [N, n_token, d_head * n_head]
+        x = ggml_nn_attention_ext(ctx, backend, q, k, v, num_heads, mask);  // [N, n_token, d_head * n_head]
 
         x = out_proj->forward(ctx, x);  // [N, n_token, model_dim]
         return {x, past_bias};
@@ -623,6 +628,7 @@ public:
     }
 
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                ggml_backend_t backend,
                                                                 struct ggml_tensor* x,
                                                                 struct ggml_tensor* past_bias                = NULL,
                                                                 struct ggml_tensor* mask                     = NULL,
@@ -632,7 +638,7 @@ public:
         auto layer_norm    = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
 
         auto normed_hidden_state = layer_norm->forward(ctx, x);
-        auto ret                 = SelfAttention->forward(ctx, normed_hidden_state, past_bias, mask, relative_position_bucket);
+        auto ret                 = SelfAttention->forward(ctx, backend, normed_hidden_state, past_bias, mask, relative_position_bucket);
         auto output              = ret.first;
         past_bias                = ret.second;
 
@@ -649,6 +655,7 @@ public:
     }
 
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                ggml_backend_t backend,
                                                                 struct ggml_tensor* x,
                                                                 struct ggml_tensor* past_bias                = NULL,
                                                                 struct ggml_tensor* mask                     = NULL,
@@ -657,7 +664,7 @@ public:
         auto layer_0 = std::dynamic_pointer_cast<T5LayerSelfAttention>(blocks["layer.0"]);
         auto layer_1 = std::dynamic_pointer_cast<T5LayerFF>(blocks["layer.1"]);
 
-        auto ret  = layer_0->forward(ctx, x, past_bias, mask, relative_position_bucket);
+        auto ret  = layer_0->forward(ctx, backend, x, past_bias, mask, relative_position_bucket);
         x         = ret.first;
         past_bias = ret.second;
         x         = layer_1->forward(ctx, x);
@@ -673,16 +680,18 @@ public:
             int64_t model_dim,
             int64_t inner_dim,
             int64_t ff_dim,
-            int64_t num_heads)
+            int64_t num_heads,
+            bool relative_attention = true)
         : num_layers(num_layers) {
         for (int i = 0; i < num_layers; i++) {
-            blocks["block." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new T5Block(model_dim, inner_dim, ff_dim, num_heads, i == 0));
+            blocks["block." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new T5Block(model_dim, inner_dim, ff_dim, num_heads, (!relative_attention || i == 0)));
         }
 
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* past_bias                = NULL,
                                 struct ggml_tensor* attention_mask           = NULL,
@@ -691,7 +700,7 @@ public:
         for (int i = 0; i < num_layers; i++) {
             auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
 
-            auto ret  = block->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
+            auto ret  = block->forward(ctx, backend, x, past_bias, attention_mask, relative_position_bucket);
             x         = ret.first;
             past_bias = ret.second;
         }
@@ -703,18 +712,34 @@ public:
     }
 };
 
+struct T5Params {
+    int64_t num_layers      = 24;
+    int64_t model_dim       = 4096;
+    int64_t ff_dim          = 10240;
+    int64_t num_heads       = 64;
+    int64_t vocab_size      = 32128;
+    bool relative_attention = true;
+};
+
 struct T5 : public GGMLBlock {
+    T5Params params;
+
 public:
-    T5(int64_t num_layers,
-       int64_t model_dim,
-       int64_t ff_dim,
-       int64_t num_heads,
-       int64_t vocab_size) {
-        blocks["encoder"] = std::shared_ptr<GGMLBlock>(new T5Stack(num_layers, model_dim, model_dim, ff_dim, num_heads));
-        blocks["shared"]  = std::shared_ptr<GGMLBlock>(new Embedding(vocab_size, model_dim));
+    T5() {}
+    T5(T5Params params)
+        : params(params) {
+        blocks["encoder"] = std::shared_ptr<GGMLBlock>(new T5Stack(params.num_layers,
+                                                                   params.model_dim,
+                                                                   params.model_dim,
+                                                                   params.ff_dim,
+                                                                   params.num_heads,
+                                                                   params.relative_attention));
+        blocks["shared"]  = std::shared_ptr<GGMLBlock>(new Embedding(params.vocab_size,
+                                                                     params.model_dim));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* past_bias                = NULL,
                                 struct ggml_tensor* attention_mask           = NULL,
@@ -725,24 +750,27 @@ public:
         auto encoder = std::dynamic_pointer_cast<T5Stack>(blocks["encoder"]);
 
         auto x = shared->forward(ctx, input_ids);
-        x      = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
+        x      = encoder->forward(ctx, backend, x, past_bias, attention_mask, relative_position_bucket);
         return x;
     }
 };
 
 struct T5Runner : public GGMLRunner {
+    T5Params params;
     T5 model;
     std::vector<int> relative_position_bucket_vec;
 
     T5Runner(ggml_backend_t backend,
-             std::map<std::string, enum ggml_type>& tensor_types,
+             bool offload_params_to_cpu,
+             const String2GGMLType& tensor_types,
              const std::string prefix,
-             int64_t num_layers = 24,
-             int64_t model_dim  = 4096,
-             int64_t ff_dim     = 10240,
-             int64_t num_heads  = 64,
-             int64_t vocab_size = 32128)
-        : GGMLRunner(backend), model(num_layers, model_dim, ff_dim, num_heads, vocab_size) {
+             bool is_umt5 = false)
+        : GGMLRunner(backend, offload_params_to_cpu) {
+        if (is_umt5) {
+            params.vocab_size         = 256384;
+            params.relative_attention = false;
+        }
+        model = T5(params);
         model.init(params_ctx, tensor_types, prefix);
     }
 
@@ -755,13 +783,14 @@ struct T5Runner : public GGMLRunner {
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* relative_position_bucket,
                                 struct ggml_tensor* attention_mask = NULL) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
 
-        auto hidden_states = model.forward(ctx, input_ids, NULL, attention_mask, relative_position_bucket);  // [N, n_token, model_dim]
+        auto hidden_states = model.forward(ctx, backend, input_ids, NULL, attention_mask, relative_position_bucket);  // [N, n_token, model_dim]
         return hidden_states;
     }
 
@@ -769,7 +798,8 @@ struct T5Runner : public GGMLRunner {
                                     struct ggml_tensor* attention_mask = NULL) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
-        input_ids = to_backend(input_ids);
+        input_ids      = to_backend(input_ids);
+        attention_mask = to_backend(attention_mask);
 
         relative_position_bucket_vec = compute_relative_position_bucket(input_ids->ne[0], input_ids->ne[0]);
 
@@ -786,7 +816,7 @@ struct T5Runner : public GGMLRunner {
                                                            input_ids->ne[0]);
         set_backend_tensor_data(relative_position_bucket, relative_position_bucket_vec.data());
 
-        struct ggml_tensor* hidden_states = forward(compute_ctx, input_ids, relative_position_bucket, attention_mask);
+        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, relative_position_bucket, attention_mask);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -876,17 +906,12 @@ struct T5Embedder {
     T5UniGramTokenizer tokenizer;
     T5Runner model;
 
-    static std::map<std::string, enum ggml_type> empty_tensor_types;
-
     T5Embedder(ggml_backend_t backend,
-               std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
-               const std::string prefix                            = "",
-               int64_t num_layers                                  = 24,
-               int64_t model_dim                                   = 4096,
-               int64_t ff_dim                                      = 10240,
-               int64_t num_heads                                   = 64,
-               int64_t vocab_size                                  = 32128)
-        : model(backend, tensor_types, prefix, num_layers, model_dim, ff_dim, num_heads, vocab_size) {
+               bool offload_params_to_cpu,
+               const String2GGMLType& tensor_types = {},
+               const std::string prefix            = "",
+               bool is_umt5                        = false)
+        : model(backend, offload_params_to_cpu, tensor_types, prefix, is_umt5), tokenizer(is_umt5) {
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@@ -948,25 +973,22 @@ struct T5Embedder {
         GGML_ASSERT(work_ctx != NULL);
 
         {
-            // cpu f16: pass
-            // cpu f32: pass
-            // cuda f16: nan
-            // cuda f32: pass
-            // cuda q8_0: nan
-            // TODO: fix cuda nan
             std::string text("a lovely cat");
-            auto tokens_and_weights     = tokenize(text, 77, true);
+            // std::string text("一只可爱的猫"); // umt5 chinease test
+            auto tokens_and_weights     = tokenize(text, 512, true);
             std::vector<int>& tokens    = std::get<0>(tokens_and_weights);
             std::vector<float>& weights = std::get<1>(tokens_and_weights);
+            std::vector<float>& masks   = std::get<2>(tokens_and_weights);
             for (auto token : tokens) {
                 printf("%d ", token);
             }
             printf("\n");
             auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+            auto attention_mask     = vector_to_ggml_tensor(work_ctx, masks);
             struct ggml_tensor* out = NULL;
 
             int t0 = ggml_time_ms();
-            model.compute(8, input_ids, NULL, &out, work_ctx);
+            model.compute(8, input_ids, attention_mask, &out, work_ctx);
             int t1 = ggml_time_ms();
 
             print_ggml_tensor(out);
@@ -975,32 +997,43 @@ struct T5Embedder {
     }
 
     static void load_from_file_and_test(const std::string& file_path) {
-        // ggml_backend_t backend    = ggml_backend_cuda_init(0);
-        ggml_backend_t backend         = ggml_backend_cpu_init();
-        ggml_type model_data_type      = GGML_TYPE_F32;
-        std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend));
-        {
-            LOG_INFO("loading from '%s'", file_path.c_str());
+        // cpu f16: pass
+        // cpu f32: pass
+        // cuda f16: pass
+        // cuda f32: pass
+        // cuda q8_0: pass
+        // ggml_backend_t backend = ggml_backend_cuda_init(0);
+        ggml_backend_t backend    = ggml_backend_cpu_init();
+        ggml_type model_data_type = GGML_TYPE_F16;
 
-            t5->alloc_params_buffer();
-            std::map<std::string, ggml_tensor*> tensors;
-            t5->get_param_tensors(tensors, "");
-
-            ModelLoader model_loader;
-            if (!model_loader.init_from_file(file_path)) {
-                LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
-                return;
-            }
-
-            bool success = model_loader.load_tensors(tensors, backend);
-
-            if (!success) {
-                LOG_ERROR("load tensors from model loader failed");
-                return;
-            }
-
-            LOG_INFO("t5 model loaded");
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
+            return;
         }
+
+        auto tensor_types = model_loader.tensor_storages_types;
+        for (auto& item : tensor_types) {
+            // LOG_DEBUG("%s %u", item.first.c_str(), item.second);
+            if (ends_with(item.first, "weight")) {
+                item.second = model_data_type;
+            }
+        }
+
+        std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend, false, tensor_types, "", true));
+
+        t5->alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> tensors;
+        t5->get_param_tensors(tensors, "");
+
+        bool success = model_loader.load_tensors(tensors);
+
+        if (!success) {
+            LOG_ERROR("load tensors from model loader failed");
+            return;
+        }
+
+        LOG_INFO("t5 model loaded");
         t5->test();
     }
 };

tae.hpp

@@ -196,16 +196,28 @@ struct TinyAutoEncoder : public GGMLRunner {
     bool decode_only = false;
 
     TinyAutoEncoder(ggml_backend_t backend,
-                    std::map<std::string, enum ggml_type>& tensor_types,
+                    bool offload_params_to_cpu,
+                    const String2GGMLType& tensor_types,
                     const std::string prefix,
                     bool decoder_only = true,
                     SDVersion version = VERSION_SD1)
         : decode_only(decoder_only),
           taesd(decoder_only, version),
-          GGMLRunner(backend) {
+          GGMLRunner(backend, offload_params_to_cpu) {
         taesd.init(params_ctx, tensor_types, prefix);
     }
 
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        taesd.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
     std::string get_desc() {
         return "taesd";
     }
@@ -226,7 +238,7 @@ struct TinyAutoEncoder : public GGMLRunner {
             return false;
         }
 
-        bool success = model_loader.load_tensors(taesd_tensors, backend, ignore_tensors);
+        bool success = model_loader.load_tensors(taesd_tensors, ignore_tensors);
 
         if (!success) {
             LOG_ERROR("load tae tensors from model loader failed");

thirdparty/darts.h

@@ -4,6 +4,7 @@
 #include <cstdio>
 #include <exception>
 #include <new>
+#include <iostream>
 
 #define DARTS_VERSION "0.32"
 
@@ -1140,9 +1141,11 @@ inline void DawgBuilder::insert(const char *key, std::size_t length,
   if (value < 0) {
     DARTS_THROW("failed to insert key: negative value");
   } else if (length == 0) {
+    std::cout << value << std::endl;
     DARTS_THROW("failed to insert key: zero-length key");
   }
 
+
   id_type id = 0;
   std::size_t key_pos = 0;
 

unet.hpp

@@ -61,6 +61,7 @@ public:
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* context,
                                 int timesteps) {
@@ -127,7 +128,7 @@ public:
             auto block     = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[transformer_name]);
             auto mix_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[time_stack_name]);
 
-            x = block->forward(ctx, x, spatial_context);  // [N, h * w, inner_dim]
+            x = block->forward(ctx, backend, x, spatial_context);  // [N, h * w, inner_dim]
 
             // in_channels == inner_dim
             auto x_mix = x;
@@ -143,7 +144,7 @@ public:
             x_mix = ggml_cont(ctx, ggml_permute(ctx, x_mix, 0, 2, 1, 3));  // b t s c -> b s t c
             x_mix = ggml_reshape_3d(ctx, x_mix, C, T, S * B);              // b s t c -> (b s) t c
 
-            x_mix = mix_block->forward(ctx, x_mix, time_context);  // [B * h * w, T, inner_dim]
+            x_mix = mix_block->forward(ctx, backend, x_mix, time_context);  // [B * h * w, T, inner_dim]
 
             x_mix = ggml_reshape_4d(ctx, x_mix, C, T, S, B);               // (b s) t c -> b s t c
             x_mix = ggml_cont(ctx, ggml_permute(ctx, x_mix, 0, 2, 1, 3));  // b s t c -> b t s c
@@ -166,7 +167,6 @@ public:
 // ldm.modules.diffusionmodules.openaimodel.UNetModel
 class UnetModelBlock : public GGMLBlock {
 protected:
-    static std::map<std::string, enum ggml_type> empty_tensor_types;
     SDVersion version = VERSION_SD1;
     // network hparams
     int in_channels                        = 4;
@@ -184,7 +184,7 @@ public:
     int model_channels  = 320;
     int adm_in_channels = 2816;  // only for VERSION_SDXL/SVD
 
-    UnetModelBlock(SDVersion version = VERSION_SD1, std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types, bool flash_attn = false)
+    UnetModelBlock(SDVersion version = VERSION_SD1, const String2GGMLType& tensor_types = {}, bool flash_attn = false)
         : version(version) {
         if (sd_version_is_sd2(version)) {
             context_dim       = 1024;
@@ -364,21 +364,23 @@ public:
 
     struct ggml_tensor* attention_layer_forward(std::string name,
                                                 struct ggml_context* ctx,
+                                                ggml_backend_t backend,
                                                 struct ggml_tensor* x,
                                                 struct ggml_tensor* context,
                                                 int timesteps) {
         if (version == VERSION_SVD) {
             auto block = std::dynamic_pointer_cast<SpatialVideoTransformer>(blocks[name]);
 
-            return block->forward(ctx, x, context, timesteps);
+            return block->forward(ctx, backend, x, context, timesteps);
         } else {
             auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
 
-            return block->forward(ctx, x, context);
+            return block->forward(ctx, backend, x, context);
         }
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* timesteps,
                                 struct ggml_tensor* context,
@@ -457,7 +459,7 @@ public:
                 h                = resblock_forward(name, ctx, h, emb, num_video_frames);  // [N, mult*model_channels, h, w]
                 if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
                     std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
-                    h                = attention_layer_forward(name, ctx, h, context, num_video_frames);  // [N, mult*model_channels, h, w]
+                    h                = attention_layer_forward(name, ctx, backend, h, context, num_video_frames);  // [N, mult*model_channels, h, w]
                 }
                 hs.push_back(h);
             }
@@ -475,9 +477,9 @@ public:
         // [N, 4*model_channels, h/8, w/8]
 
         // middle_block
-        h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
-        h = attention_layer_forward("middle_block.1", ctx, h, context, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
-        h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames);                      // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, backend, h, context, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);                      // [N, 4*model_channels, h/8, w/8]
 
         if (controls.size() > 0) {
             auto cs = ggml_scale_inplace(ctx, controls[controls.size() - 1], control_strength);
@@ -508,7 +510,7 @@ public:
                 if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
                     std::string name = "output_blocks." + std::to_string(output_block_idx) + ".1";
 
-                    h = attention_layer_forward(name, ctx, h, context, num_video_frames);
+                    h = attention_layer_forward(name, ctx, backend, h, context, num_video_frames);
 
                     up_sample_idx++;
                 }
@@ -539,14 +541,27 @@ struct UNetModelRunner : public GGMLRunner {
     UnetModelBlock unet;
 
     UNetModelRunner(ggml_backend_t backend,
-                    std::map<std::string, enum ggml_type>& tensor_types,
+                    bool offload_params_to_cpu,
+                    const String2GGMLType& tensor_types,
                     const std::string prefix,
                     SDVersion version = VERSION_SD1,
                     bool flash_attn   = false)
-        : GGMLRunner(backend), unet(version, tensor_types, flash_attn) {
+        : GGMLRunner(backend, offload_params_to_cpu), unet(version, tensor_types, flash_attn) {
         unet.init(params_ctx, tensor_types, prefix);
     }
 
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        unet.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                LOG_DEBUG("block %s", block->get_desc().c_str());
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
     std::string get_desc() {
         return "unet";
     }
@@ -580,6 +595,7 @@ struct UNetModelRunner : public GGMLRunner {
         }
 
         struct ggml_tensor* out = unet.forward(compute_ctx,
+                                               runtime_backend,
                                                x,
                                                timesteps,
                                                context,

upscaler.cpp

@@ -9,19 +9,23 @@ struct UpscalerGGML {
     std::shared_ptr<ESRGAN> esrgan_upscaler;
     std::string esrgan_path;
     int n_threads;
+    bool direct = false;
 
-    UpscalerGGML(int n_threads)
-        : n_threads(n_threads) {
+    UpscalerGGML(int n_threads,
+                 bool direct = false)
+        : n_threads(n_threads),
+          direct(direct) {
     }
 
-    bool load_from_file(const std::string& esrgan_path) {
+    bool load_from_file(const std::string& esrgan_path,
+                        bool offload_params_to_cpu) {
+        ggml_log_set(ggml_log_callback_default, nullptr);
 #ifdef SD_USE_CUDA
         LOG_DEBUG("Using CUDA backend");
         backend = ggml_backend_cuda_init(0);
 #endif
 #ifdef SD_USE_METAL
         LOG_DEBUG("Using Metal backend");
-        ggml_log_set(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
 #ifdef SD_USE_VULKAN
@@ -46,7 +50,10 @@ struct UpscalerGGML {
             backend = ggml_backend_cpu_init();
         }
         LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
-        esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_loader.tensor_storages_types);
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend, offload_params_to_cpu, model_loader.tensor_storages_types);
+        if (direct) {
+            esrgan_upscaler->enable_conv2d_direct();
+        }
         if (!esrgan_upscaler->load_from_file(esrgan_path)) {
             return false;
         }
@@ -62,8 +69,7 @@ struct UpscalerGGML {
                  input_image.width, input_image.height, output_width, output_height);
 
         struct ggml_init_params params;
-        params.mem_size = output_width * output_height * 3 * sizeof(float) * 2;
-        params.mem_size += 2 * ggml_tensor_overhead();
+        params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
         params.mem_buffer = NULL;
         params.no_alloc   = false;
 
@@ -73,9 +79,9 @@ struct UpscalerGGML {
             LOG_ERROR("ggml_init() failed");
             return upscaled_image;
         }
-        LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
+        // LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
         ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
-        sd_image_to_tensor(input_image.data, input_image_tensor);
+        sd_image_to_tensor(input_image, input_image_tensor);
 
         ggml_tensor* upscaled = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, output_width, output_height, 3, 1);
         auto on_tiling        = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
@@ -104,6 +110,8 @@ struct upscaler_ctx_t {
 };
 
 upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
+                                 bool offload_params_to_cpu,
+                                 bool direct,
                                  int n_threads) {
     upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
     if (upscaler_ctx == NULL) {
@@ -111,12 +119,12 @@ upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
     }
     std::string esrgan_path(esrgan_path_c_str);
 
-    upscaler_ctx->upscaler = new UpscalerGGML(n_threads);
+    upscaler_ctx->upscaler = new UpscalerGGML(n_threads, direct);
     if (upscaler_ctx->upscaler == NULL) {
         return NULL;
     }
 
-    if (!upscaler_ctx->upscaler->load_from_file(esrgan_path)) {
+    if (!upscaler_ctx->upscaler->load_from_file(esrgan_path, offload_params_to_cpu)) {
         delete upscaler_ctx->upscaler;
         upscaler_ctx->upscaler = NULL;
         free(upscaler_ctx);

util.cpp

@@ -72,6 +72,17 @@ std::string format(const char* fmt, ...) {
     return std::string(buf.data(), size);
 }
 
+int round_up_to(int value, int base) {
+    if (base <= 0) {
+        return value;
+    }
+    if (value % base == 0) {
+        return value;
+    } else {
+        return ((value / base) + 1) * base;
+    }
+}
+
 #ifdef _WIN32  // code for windows
 #include <windows.h>
 
@@ -99,56 +110,6 @@ std::string get_full_path(const std::string& dir, const std::string& filename) {
     }
 }
 
-std::vector<std::string> get_files_from_dir(const std::string& dir) {
-    std::vector<std::string> files;
-
-    WIN32_FIND_DATA findFileData;
-    HANDLE hFind;
-
-    char currentDirectory[MAX_PATH];
-    GetCurrentDirectory(MAX_PATH, currentDirectory);
-
-    char directoryPath[MAX_PATH];  // this is absolute path
-    sprintf(directoryPath, "%s\\%s\\*", currentDirectory, dir.c_str());
-
-    // Find the first file in the directory
-    hFind               = FindFirstFile(directoryPath, &findFileData);
-    bool isAbsolutePath = false;
-    // Check if the directory was found
-    if (hFind == INVALID_HANDLE_VALUE) {
-        printf("Unable to find directory. Try with original path \n");
-
-        char directoryPathAbsolute[MAX_PATH];
-        sprintf(directoryPathAbsolute, "%s*", dir.c_str());
-
-        hFind          = FindFirstFile(directoryPathAbsolute, &findFileData);
-        isAbsolutePath = true;
-        if (hFind == INVALID_HANDLE_VALUE) {
-            printf("Absolute path was also wrong.\n");
-            return files;
-        }
-    }
-
-    // Loop through all files in the directory
-    do {
-        // Check if the found file is a regular file (not a directory)
-        if (!(findFileData.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY)) {
-            if (isAbsolutePath) {
-                files.push_back(dir + "\\" + std::string(findFileData.cFileName));
-            } else {
-                files.push_back(std::string(currentDirectory) + "\\" + dir + "\\" + std::string(findFileData.cFileName));
-            }
-        }
-    } while (FindNextFile(hFind, &findFileData) != 0);
-
-    // Close the handle
-    FindClose(hFind);
-
-    sort(files.begin(), files.end());
-
-    return files;
-}
-
 #else  // Unix
 #include <dirent.h>
 #include <sys/stat.h>
@@ -183,27 +144,6 @@ std::string get_full_path(const std::string& dir, const std::string& filename) {
     return "";
 }
 
-std::vector<std::string> get_files_from_dir(const std::string& dir) {
-    std::vector<std::string> files;
-
-    DIR* dp = opendir(dir.c_str());
-
-    if (dp != nullptr) {
-        struct dirent* entry;
-
-        while ((entry = readdir(dp)) != nullptr) {
-            std::string fname = dir + "/" + entry->d_name;
-            if (!is_directory(fname))
-                files.push_back(fname);
-        }
-        closedir(dp);
-    }
-
-    sort(files.begin(), files.end());
-
-    return files;
-}
-
 #endif
 
 // get_num_physical_cores is copy from
@@ -290,7 +230,7 @@ std::string path_join(const std::string& p1, const std::string& p2) {
     return p1 + "/" + p2;
 }
 
-std::vector<std::string> splitString(const std::string& str, char delimiter) {
+std::vector<std::string> split_string(const std::string& str, char delimiter) {
     std::vector<std::string> result;
     size_t start = 0;
     size_t end   = str.find(delimiter);
@@ -307,39 +247,6 @@ std::vector<std::string> splitString(const std::string& str, char delimiter) {
     return result;
 }
 
-sd_image_t* preprocess_id_image(sd_image_t* img) {
-    int shortest_edge   = 224;
-    int size            = shortest_edge;
-    sd_image_t* resized = NULL;
-    uint32_t w          = img->width;
-    uint32_t h          = img->height;
-    uint32_t c          = img->channel;
-
-    // 1. do resize using stb_resize functions
-
-    unsigned char* buf = (unsigned char*)malloc(sizeof(unsigned char) * 3 * size * size);
-    if (!stbir_resize_uint8(img->data, w, h, 0,
-                            buf, size, size, 0,
-                            c)) {
-        fprintf(stderr, "%s: resize operation failed \n ", __func__);
-        return resized;
-    }
-
-    // 2. do center crop (likely unnecessary due to step 1)
-
-    // 3. do rescale
-
-    // 4. do normalize
-
-    // 3 and 4 will need to be done in float format.
-
-    resized = new sd_image_t{(uint32_t)shortest_edge,
-                             (uint32_t)shortest_edge,
-                             3,
-                             buf};
-    return resized;
-}
-
 void pretty_progress(int step, int steps, float time) {
     if (sd_progress_cb) {
         sd_progress_cb(step, steps, time, sd_progress_cb_data);
@@ -403,7 +310,10 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
     if (written >= 0 && written < LOG_BUFFER_SIZE) {
         vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
     }
-    strncat(log_buffer, "\n", LOG_BUFFER_SIZE - strlen(log_buffer));
+    size_t len = strlen(log_buffer);
+    if (log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
 
     if (sd_log_cb) {
         sd_log_cb(level, log_buffer, sd_log_cb_data);

util.h

@@ -18,18 +18,15 @@ std::string format(const char* fmt, ...);
 
 void replace_all_chars(std::string& str, char target, char replacement);
 
+int round_up_to(int value, int base);
+
 bool file_exists(const std::string& filename);
 bool is_directory(const std::string& path);
 std::string get_full_path(const std::string& dir, const std::string& filename);
 
-std::vector<std::string> get_files_from_dir(const std::string& dir);
-
 std::u32string utf8_to_utf32(const std::string& utf8_str);
 std::string utf32_to_utf8(const std::u32string& utf32_str);
 std::u32string unicode_value_to_utf32(int unicode_value);
-
-sd_image_t* preprocess_id_image(sd_image_t* img);
-
 // std::string sd_basename(const std::string& path);
 
 typedef struct {
@@ -48,7 +45,7 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
 sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size);
 
 std::string path_join(const std::string& p1, const std::string& p2);
-std::vector<std::string> splitString(const std::string& str, char delimiter);
+std::vector<std::string> split_string(const std::string& str, char delimiter);
 void pretty_progress(int step, int steps, float time);
 
 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);

vae.hpp

@@ -163,8 +163,8 @@ public:
 
 class VideoResnetBlock : public ResnetBlock {
 protected:
-    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
-        enum ggml_type wtype = (tensor_types.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type wtype = get_type(prefix + "mix_factor", tensor_types, GGML_TYPE_F32);
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 
@@ -520,20 +520,44 @@ public:
     }
 };
 
-struct AutoEncoderKL : public GGMLRunner {
+struct VAE : public GGMLRunner {
+    VAE(ggml_backend_t backend, bool offload_params_to_cpu)
+        : GGMLRunner(backend, offload_params_to_cpu) {}
+    virtual void compute(const int n_threads,
+                         struct ggml_tensor* z,
+                         bool decode_graph,
+                         struct ggml_tensor** output,
+                         struct ggml_context* output_ctx)                                                         = 0;
+    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
+    virtual void enable_conv2d_direct(){};
+};
+
+struct AutoEncoderKL : public VAE {
     bool decode_only = true;
     AutoencodingEngine ae;
 
     AutoEncoderKL(ggml_backend_t backend,
-                  std::map<std::string, enum ggml_type>& tensor_types,
+                  bool offload_params_to_cpu,
+                  const String2GGMLType& tensor_types,
                   const std::string prefix,
                   bool decode_only       = false,
                   bool use_video_decoder = false,
                   SDVersion version      = VERSION_SD1)
-        : decode_only(decode_only), ae(decode_only, use_video_decoder, version), GGMLRunner(backend) {
+        : decode_only(decode_only), ae(decode_only, use_video_decoder, version), VAE(backend, offload_params_to_cpu) {
         ae.init(params_ctx, tensor_types, prefix);
     }
 
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        ae.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
     std::string get_desc() {
         return "vae";
     }
@@ -564,7 +588,7 @@ struct AutoEncoderKL : public GGMLRunner {
         };
         // ggml_set_f32(z, 0.5f);
         // print_ggml_tensor(z);
-        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
     }
 
     void test() {