sync: update ggml to fix large image generation with SYCL backend (#380)

* turn off fast-math on host in SYCL backend

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* update ggml for sync some sycl ops

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

* update sycl readme and ggml

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

---------

Signed-off-by: zhentaoyu <zhentao.yu@intel.com>

.github/workflows/build.yml

@@ -148,6 +148,9 @@ jobs:
   windows-latest-cmake:
     runs-on: windows-2019
 
+    env:
+      VULKAN_VERSION: 1.3.261.1
+
     strategy:
       matrix:
         include:
@@ -163,6 +166,8 @@ jobs:
             defines: "-DSD_CUBLAS=ON -DSD_BUILD_SHARED_LIBS=ON"
           - 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'
+            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
     steps:
       - name: Clone
         id: checkout
@@ -192,6 +197,14 @@ jobs:
         uses: urkle/action-get-ninja@v1
         with:
           version: 1.11.1
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'vulkan' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
 
       - name: Build
         id: cmake_build

CMakeLists.txt

@@ -27,6 +27,7 @@ option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
 option(SD_CUBLAS                     "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
+option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
 option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
@@ -34,19 +35,25 @@ option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
 if(SD_CUBLAS)
-    message("Use CUBLAS as backend stable-diffusion")
+    message("-- Use CUBLAS as backend stable-diffusion")
     set(GGML_CUDA ON)
     add_definitions(-DSD_USE_CUBLAS)
 endif()
 
 if(SD_METAL)
-    message("Use Metal as backend stable-diffusion")
+    message("-- Use Metal as backend stable-diffusion")
     set(GGML_METAL ON)
     add_definitions(-DSD_USE_METAL)
 endif()
 
+if (SD_VULKAN)
+    message("-- Use Vulkan as backend stable-diffusion")
+    set(GGML_VULKAN ON)
+    add_definitions(-DSD_USE_VULKAN)
+endif ()
+
 if (SD_HIPBLAS)
-    message("Use HIPBLAS as backend stable-diffusion")
+    message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIPBLAS ON)
     add_definitions(-DSD_USE_CUBLAS)
     if(SD_FAST_SOFTMAX)
@@ -54,14 +61,8 @@ if (SD_HIPBLAS)
     endif()
 endif ()
 
-if(SD_SYCL)
-    message("Use SYCL as backend stable-diffusion")
-    set(GGML_SYCL ON)
-    add_definitions(-DSD_USE_SYCL)
-endif()
-
 if(SD_FLASH_ATTN)
-    message("Use Flash Attention for memory optimization")
+    message("-- Use Flash Attention for memory optimization")
     add_definitions(-DSD_USE_FLASH_ATTENTION)
 endif()
 
@@ -75,7 +76,7 @@ file(GLOB SD_LIB_SOURCES
 
 # we can get only one share lib
 if(SD_BUILD_SHARED_LIBS)
-    message("Build shared library")
+    message("-- Build shared library")
     message(${SD_LIB_SOURCES})
     set(BUILD_SHARED_LIBS OFF)
     add_library(${SD_LIB} SHARED ${SD_LIB_SOURCES})
@@ -83,11 +84,25 @@ if(SD_BUILD_SHARED_LIBS)
     target_compile_definitions(${SD_LIB} PRIVATE -DSD_BUILD_DLL)
     set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 else()
-    message("Build static library")
+    message("-- Build static library")
     set(BUILD_SHARED_LIBS OFF)
     add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()
 
+if(SD_SYCL)
+    message("-- Use SYCL as backend stable-diffusion")
+    set(GGML_SYCL ON)
+    add_definitions(-DSD_USE_SYCL)
+    # disable fast-math on host, see:
+    # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
+    if (WIN32)
+        set(SYCL_COMPILE_OPTIONS /fp:precise)
+    else()
+        set(SYCL_COMPILE_OPTIONS -fp-model=precise)
+    endif()
+    message("-- Turn off fast-math for host in SYCL backend")
+    target_compile_options(${SD_LIB} PRIVATE ${SYCL_COMPILE_OPTIONS})
+endif()
 
 set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
@@ -95,7 +110,10 @@ set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 add_definitions(-DGGML_MAX_NAME=128)
 
 # deps
-add_subdirectory(ggml)
+# Only add ggml if it hasn't been added yet
+if (NOT TARGET ggml)
+    add_subdirectory(ggml)
+endif()
 
 add_subdirectory(thirdparty)
 

README.md

@@ -4,7 +4,7 @@
 
 # stable-diffusion.cpp
 
-Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in pure C/C++
+Inference of Stable Diffusion and Flux in pure C/C++
 
 ## Features
 
@@ -21,7 +21,7 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - Accelerated memory-efficient CPU inference
     - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image, enabling Flash Attention just requires ~1.8GB.
 - AVX, AVX2 and AVX512 support for x86 architectures
-- Full CUDA, Metal and SYCL backend for GPU acceleration.
+- Full CUDA, Metal, Vulkan and SYCL backend for GPU acceleration.
 - 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 (only cpu for now)
@@ -142,6 +142,15 @@ cmake .. -DSD_METAL=ON
 cmake --build . --config Release
 ```
 
+##### Using Vulkan
+
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+
+```
+cmake .. -DSD_VULKAN=ON
+cmake --build . --config Release
+```
+
 ##### Using SYCL
 
 Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
@@ -163,14 +172,12 @@ Example of text2img by using SYCL backend:
 
 - download `stable-diffusion` model weight, refer to [download-weight](#download-weights).
 
-- run `./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 30 --sampling-method euler  -H 512 -W 512 --seed 42 -p "fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details, 4k resolution"`
+- run `./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 30 --sampling-method euler  -H 1024 -W 1024 --seed 42 -p "fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details, 4k resolution"`
 
 <p align="center">
   <img src="./assets/sycl_sd3_output.png" width="360x">
 </p>
 
-> [!NOTE]
-> Try to set smaller image height and width (for example, `-H 512 -W 512`) if you meet `Provided range is out of integer limits. Pass '-fno-sycl-id-queries-fit-in-int' to disable range check.`
 
 
 ##### Using Flash Attention
@@ -192,7 +199,10 @@ arguments:
   -M, --mode [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)
   -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
-  -m, --model [MODEL]                path to model
+  -m, --model [MODEL]                path to full model
+  --diffusion-model                  path to the standalone diffusion model
+  --clip_l                           path to the clip-l text encoder
+  --t5xxl                            path to the 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
@@ -217,16 +227,18 @@ arguments:
                                      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, lcm}
+  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm}
                                      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, ays} Denoiser sigma schedule (default: discrete)
+  --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)
                                      <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   --vae-tiling                       process vae in tiles to reduce memory usage
+  --vae-on-cpu                       keep vae in cpu (for low vram)
+  --clip-on-cpu                      keep clip in cpu (for low vram).
   --control-net-cpu                  keep controlnet in cpu (for low vram)
   --canny                            apply canny preprocessor (edge detection)
   --color                            Colors the logging tags according to level
@@ -291,6 +303,10 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 
 [![Contributors](https://contrib.rocks/image?repo=leejet/stable-diffusion.cpp)](https://github.com/leejet/stable-diffusion.cpp/graphs/contributors)
 
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=leejet/stable-diffusion.cpp&type=Date)](https://star-history.com/#leejet/stable-diffusion.cpp&Date)
+
 ## References
 
 - [ggml](https://github.com/ggerganov/ggml)

clip.hpp

@@ -76,7 +76,7 @@ public:
     const std::string UNK_TOKEN = "<|endoftext|>";
     const std::string BOS_TOKEN = "<|startoftext|>";
     const std::string EOS_TOKEN = "<|endoftext|>";
-    const std::string PAD_TOEKN = "<|endoftext|>";
+    const std::string PAD_TOKEN = "<|endoftext|>";
 
     const int UNK_TOKEN_ID = 49407;
     const int BOS_TOKEN_ID = 49406;
@@ -922,4 +922,4 @@ struct CLIPTextModelRunner : public GGMLRunner {
     }
 };
 
-#endif  // __CLIP_HPP__
+#endif  // __CLIP_HPP__

conditioner.hpp

@@ -10,8 +10,8 @@ struct SDCondition {
     struct ggml_tensor* c_concat    = NULL;
 
     SDCondition() = default;
-    SDCondition(struct ggml_tensor* c_crossattn, struct ggml_tensor* c_vector, struct ggml_tensor* c_concat) :
-    c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat) {}
+    SDCondition(struct ggml_tensor* c_crossattn, struct ggml_tensor* c_vector, struct ggml_tensor* c_concat)
+        : c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat) {}
 };
 
 struct Conditioner {
@@ -978,7 +978,6 @@ struct SD3CLIPEmbedder : public Conditioner {
     }
 };
 
-
 struct FluxCLIPEmbedder : public Conditioner {
     ggml_type wtype;
     CLIPTokenizer clip_l_tokenizer;
@@ -987,8 +986,8 @@ struct FluxCLIPEmbedder : public Conditioner {
     std::shared_ptr<T5Runner> t5;
 
     FluxCLIPEmbedder(ggml_backend_t backend,
-                    ggml_type wtype,
-                    int clip_skip = -1)
+                     ggml_type wtype,
+                     int clip_skip = -1)
         : wtype(wtype) {
         if (clip_skip <= 0) {
             clip_skip = 2;
@@ -1085,10 +1084,10 @@ struct FluxCLIPEmbedder : public Conditioner {
         auto& t5_tokens      = token_and_weights[1].first;
         auto& t5_weights     = token_and_weights[1].second;
 
-        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,]
+        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,]
         std::vector<float> hidden_states_vec;
 
         size_t chunk_len   = 256;

denoiser.hpp

@@ -2,6 +2,7 @@
 #define __DENOISER_HPP__
 
 #include "ggml_extend.hpp"
+#include "gits_noise.inl"
 
 /*================================================= CompVisDenoiser ==================================================*/
 
@@ -41,91 +42,114 @@ struct DiscreteSchedule : SigmaSchedule {
     }
 };
 
+struct ExponentialSchedule : SigmaSchedule {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+        std::vector<float> sigmas;
+
+        // Calculate step size
+        float log_sigma_min = std::log(sigma_min);
+        float log_sigma_max = std::log(sigma_max);
+        float step = (log_sigma_max - log_sigma_min) / (n - 1);
+
+        // Fill sigmas with exponential values
+        for (uint32_t i = 0; i < n; ++i) {
+            float sigma = std::exp(log_sigma_max - step * i);
+            sigmas.push_back(sigma);
+        }
+
+        sigmas.push_back(0.0f);
+
+        return sigmas;
+    }
+};
+
+/* interp and linear_interp adapted from dpilger26's NumCpp library:
+ * https://github.com/dpilger26/NumCpp/tree/5e40aab74d14e257d65d3dc385c9ff9e2120c60e */
+constexpr double interp(double left, double right, double perc) noexcept {
+    return (left * (1. - perc)) + (right * perc);
+}
+
+/* This will make the assumption that the reference x and y values are
+ * already sorted in ascending order because they are being generated as
+ * such in the calling function */
+std::vector<double> linear_interp(std::vector<float> new_x,
+                                  const std::vector<float> ref_x,
+                                  const std::vector<float> ref_y) {
+    const size_t len_x = new_x.size();
+    size_t i           = 0;
+    size_t j           = 0;
+    std::vector<double> new_y(len_x);
+
+    if (ref_x.size() != ref_y.size()) {
+        LOG_ERROR("Linear Interpolation Failed: length mismatch");
+        return new_y;
+    }
+
+    /* Adjusted bounds checking to ensure new_x is within ref_x range */
+    if (new_x[0] < ref_x[0]) {
+        new_x[0] = ref_x[0];
+    }
+    if (new_x.back() > ref_x.back()) {
+        new_x.back() = ref_x.back();
+    }
+
+    while (i < len_x) {
+        if ((ref_x[j] > new_x[i]) || (new_x[i] > ref_x[j + 1])) {
+            j++;
+            continue;
+        }
+
+        const double perc = static_cast<double>(new_x[i] - ref_x[j]) / static_cast<double>(ref_x[j + 1] - ref_x[j]);
+
+        new_y[i] = interp(ref_y[j], ref_y[j + 1], perc);
+        i++;
+    }
+
+    return new_y;
+}
+
+std::vector<float> linear_space(const float start, const float end, const size_t num_points) {
+    std::vector<float> result(num_points);
+    const float inc = (end - start) / (static_cast<float>(num_points - 1));
+
+    if (num_points > 0) {
+        result[0] = start;
+
+        for (size_t i = 1; i < num_points; i++) {
+            result[i] = result[i - 1] + inc;
+        }
+    }
+
+    return result;
+}
+
+std::vector<float> log_linear_interpolation(std::vector<float> sigma_in,
+                                            const size_t new_len) {
+    const size_t s_len        = sigma_in.size();
+    std::vector<float> x_vals = linear_space(0.f, 1.f, s_len);
+    std::vector<float> y_vals(s_len);
+
+    /* Reverses the input array to be ascending instead of descending,
+     * also hits it with a log, it is log-linear interpolation after all */
+    for (size_t i = 0; i < s_len; i++) {
+        y_vals[i] = std::log(sigma_in[s_len - i - 1]);
+    }
+
+    std::vector<float> new_x_vals  = linear_space(0.f, 1.f, new_len);
+    std::vector<double> new_y_vals = linear_interp(new_x_vals, x_vals, y_vals);
+    std::vector<float> results(new_len);
+
+    for (size_t i = 0; i < new_len; i++) {
+        results[i] = static_cast<float>(std::exp(new_y_vals[new_len - i - 1]));
+    }
+
+    return results;
+}
+
 /*
 https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/howto.html
 */
 struct AYSSchedule : SigmaSchedule {
-    /* interp and linear_interp adapted from dpilger26's NumCpp library:
-     * https://github.com/dpilger26/NumCpp/tree/5e40aab74d14e257d65d3dc385c9ff9e2120c60e */
-    constexpr double interp(double left, double right, double perc) noexcept {
-        return (left * (1. - perc)) + (right * perc);
-    }
-
-    /* This will make the assumption that the reference x and y values are
-     * already sorted in ascending order because they are being generated as
-     * such in the calling function */
-    std::vector<double> linear_interp(std::vector<float> new_x,
-                                      const std::vector<float> ref_x,
-                                      const std::vector<float> ref_y) {
-        const size_t len_x = new_x.size();
-        size_t i           = 0;
-        size_t j           = 0;
-        std::vector<double> new_y(len_x);
-
-        if (ref_x.size() != ref_y.size()) {
-            LOG_ERROR("Linear Interoplation Failed: length mismatch");
-            return new_y;
-        }
-
-        /* serves as the bounds checking for the below while loop */
-        if ((new_x[0] < ref_x[0]) || (new_x[new_x.size() - 1] > ref_x[ref_x.size() - 1])) {
-            LOG_ERROR("Linear Interpolation Failed: bad bounds");
-            return new_y;
-        }
-
-        while (i < len_x) {
-            if ((ref_x[j] > new_x[i]) || (new_x[i] > ref_x[j + 1])) {
-                j++;
-                continue;
-            }
-
-            const double perc = static_cast<double>(new_x[i] - ref_x[j]) / static_cast<double>(ref_x[j + 1] - ref_x[j]);
-
-            new_y[i] = interp(ref_y[j], ref_y[j + 1], perc);
-            i++;
-        }
-
-        return new_y;
-    }
-
-    std::vector<float> linear_space(const float start, const float end, const size_t num_points) {
-        std::vector<float> result(num_points);
-        const float inc = (end - start) / (static_cast<float>(num_points - 1));
-
-        if (num_points > 0) {
-            result[0] = start;
-
-            for (size_t i = 1; i < num_points; i++) {
-                result[i] = result[i - 1] + inc;
-            }
-        }
-
-        return result;
-    }
-
-    std::vector<float> log_linear_interpolation(std::vector<float> sigma_in,
-                                                const size_t new_len) {
-        const size_t s_len        = sigma_in.size();
-        std::vector<float> x_vals = linear_space(0.f, 1.f, s_len);
-        std::vector<float> y_vals(s_len);
-
-        /* Reverses the input array to be ascending instead of descending,
-         * also hits it with a log, it is log-linear interpolation after all */
-        for (size_t i = 0; i < s_len; i++) {
-            y_vals[i] = std::log(sigma_in[s_len - i - 1]);
-        }
-
-        std::vector<float> new_x_vals  = linear_space(0.f, 1.f, new_len);
-        std::vector<double> new_y_vals = linear_interp(new_x_vals, x_vals, y_vals);
-        std::vector<float> results(new_len);
-
-        for (size_t i = 0; i < new_len; i++) {
-            results[i] = static_cast<float>(std::exp(new_y_vals[new_len - i - 1]));
-        }
-
-        return results;
-    }
-
     std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
         const std::vector<float> noise_levels[] = {
             /* SD1.5 */
@@ -179,6 +203,38 @@ struct AYSSchedule : SigmaSchedule {
     }
 };
 
+/*
+ * GITS Scheduler: https://github.com/zju-pi/diff-sampler/tree/main/gits-main
+*/
+struct GITSSchedule : SigmaSchedule {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+        if (sigma_max <= 0.0f) {
+            return std::vector<float>{};
+        }
+
+        std::vector<float> sigmas;
+
+        // Assume coeff is provided (replace 1.20 with your dynamic coeff)
+        float coeff = 1.20f;  // Default coefficient
+        // Normalize coeff to the closest value in the array (0.80 to 1.50)
+        coeff = std::round(coeff * 20.0f) / 20.0f;  // Round to the nearest 0.05
+        // Calculate the index based on the coefficient
+        int index = static_cast<int>((coeff - 0.80f) / 0.05f);
+        // Ensure the index is within bounds
+        index = std::max(0, std::min(index, static_cast<int>(GITS_NOISE.size() - 1)));
+        const std::vector<std::vector<float>>& selected_noise = *GITS_NOISE[index];
+
+        if (n <= 20) {
+            sigmas = (selected_noise)[n - 2];
+        } else {
+            sigmas = log_linear_interpolation(selected_noise.back(), n + 1);
+        }
+
+        sigmas[n] = 0.0f;
+        return sigmas;
+    }
+};
+
 struct KarrasSchedule : SigmaSchedule {
     std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
         // These *COULD* be function arguments here,
@@ -351,7 +407,6 @@ struct DiscreteFlowDenoiser : public Denoiser {
     }
 };
 
-
 float flux_time_shift(float mu, float sigma, float t) {
     return std::exp(mu) / (std::exp(mu) + std::pow((1.0 / t - 1.0), sigma));
 }
@@ -369,7 +424,7 @@ struct FluxFlowDenoiser : public Denoiser {
     void set_parameters(float shift = 1.15f) {
         this->shift = shift;
         for (int i = 1; i < TIMESTEPS + 1; i++) {
-            sigmas[i - 1] = t_to_sigma(i/TIMESTEPS * TIMESTEPS);
+            sigmas[i - 1] = t_to_sigma(i / TIMESTEPS * TIMESTEPS);
         }
     }
 
@@ -766,6 +821,158 @@ static void sample_k_diffusion(sample_method_t method,
                 }
             }
         } break;
+        case IPNDM:  // iPNDM sampler from https://github.com/zju-pi/diff-sampler/tree/main/diff-solvers-main
+        {
+            int max_order = 4;
+            ggml_tensor* x_next = x;
+            std::vector<ggml_tensor*> buffer_model;
+
+            for (int i = 0; i < steps; i++) {
+                float sigma = sigmas[i];
+                float sigma_next = sigmas[i + 1];
+
+                ggml_tensor* x_cur = x_next;
+                float* vec_x_cur = (float*)x_cur->data;
+                float* vec_x_next = (float*)x_next->data;
+
+                // Denoising step
+                ggml_tensor* denoised = model(x_cur, sigma, i + 1);
+                float* vec_denoised = (float*)denoised->data;
+                // d_cur = (x_cur - denoised) / sigma
+                struct ggml_tensor* d_cur = ggml_dup_tensor(work_ctx, x_cur);
+                float* vec_d_cur = (float*)d_cur->data;
+
+                for (int j = 0; j < ggml_nelements(d_cur); j++) {
+                    vec_d_cur[j] = (vec_x_cur[j] - vec_denoised[j]) / sigma;
+                }
+
+                int order = std::min(max_order, i + 1);
+
+                // Calculate vec_x_next based on the order
+                switch (order) {
+                    case 1:  // First Euler step
+                        for (int j = 0; j < ggml_nelements(x_next); j++) {
+                            vec_x_next[j] = vec_x_cur[j] + (sigma_next - sigma) * vec_d_cur[j];
+                        }
+                        break;
+
+                    case 2:  // Use one history point
+                        {
+                            float* vec_d_prev1 = (float*)buffer_model.back()->data;
+                            for (int j = 0; j < ggml_nelements(x_next); j++) {
+                                vec_x_next[j] = vec_x_cur[j] + (sigma_next - sigma) * (3 * vec_d_cur[j] - vec_d_prev1[j]) / 2;
+                            }
+                        }
+                        break;
+
+                    case 3:  // Use two history points
+                        {
+                            float* vec_d_prev1 = (float*)buffer_model.back()->data;
+                            float* vec_d_prev2 = (float*)buffer_model[buffer_model.size() - 2]->data;
+                            for (int j = 0; j < ggml_nelements(x_next); j++) {
+                                vec_x_next[j] = vec_x_cur[j] + (sigma_next - sigma) * (23 * vec_d_cur[j] - 16 * vec_d_prev1[j] + 5 * vec_d_prev2[j]) / 12;
+                            }
+                        }
+                        break;
+
+                    case 4:  // Use three history points
+                        {
+                            float* vec_d_prev1 = (float*)buffer_model.back()->data;
+                            float* vec_d_prev2 = (float*)buffer_model[buffer_model.size() - 2]->data;
+                            float* vec_d_prev3 = (float*)buffer_model[buffer_model.size() - 3]->data;
+                            for (int j = 0; j < ggml_nelements(x_next); j++) {
+                                vec_x_next[j] = vec_x_cur[j] + (sigma_next - sigma) * (55 * vec_d_cur[j] - 59 * vec_d_prev1[j] + 37 * vec_d_prev2[j] - 9 * vec_d_prev3[j]) / 24;
+                            }
+                        }
+                        break;
+                }
+
+                // Manage buffer_model
+                if (buffer_model.size() == max_order - 1) {
+                    // Shift elements to the left
+                    for (int k = 0; k < max_order - 2; k++) {
+                        buffer_model[k] = buffer_model[k + 1];
+                    }
+                    buffer_model.back() = d_cur;  // Replace the last element with d_cur
+                } else {
+                    buffer_model.push_back(d_cur);
+                }
+            }
+        } break;
+        case IPNDM_V:  // iPNDM_v sampler from https://github.com/zju-pi/diff-sampler/tree/main/diff-solvers-main
+        {
+            int max_order = 4;
+            std::vector<ggml_tensor*> buffer_model;
+            ggml_tensor* x_next = x;
+
+            for (int i = 0; i < steps; i++) {
+                float sigma = sigmas[i];
+                float t_next = sigmas[i + 1];
+
+                // Denoising step
+                ggml_tensor* denoised = model(x, sigma, i + 1);
+                float* vec_denoised = (float*)denoised->data;
+                struct ggml_tensor* d_cur = ggml_dup_tensor(work_ctx, x);
+                float* vec_d_cur = (float*)d_cur->data;
+                float* vec_x = (float*)x->data;
+
+                // d_cur = (x - denoised) / sigma
+                for (int j = 0; j < ggml_nelements(d_cur); j++) {
+                    vec_d_cur[j] = (vec_x[j] - vec_denoised[j]) / sigma;
+                }
+
+                int order = std::min(max_order, i + 1);
+                float h_n = t_next - sigma;
+                float h_n_1 = (i > 0) ? (sigma - sigmas[i - 1]) : h_n;
+
+                switch (order) {
+                    case 1:  // First Euler step 
+                        for (int j = 0; j < ggml_nelements(x_next); j++) {
+                            vec_x[j] += vec_d_cur[j] * h_n;
+                        }
+                        break;
+
+                    case 2: {
+                        float* vec_d_prev1 = (float*)buffer_model.back()->data;
+                        for (int j = 0; j < ggml_nelements(x_next); j++) {
+                            vec_x[j] += h_n * ((2 + (h_n / h_n_1)) * vec_d_cur[j] - (h_n / h_n_1) * vec_d_prev1[j]) / 2;
+                        }
+                        break;
+                    }
+
+                    case 3: {
+                        float h_n_2 = (i > 1) ? (sigmas[i - 1] - sigmas[i - 2]) : h_n_1;
+                        float* vec_d_prev1 = (float*)buffer_model.back()->data;
+                        float* vec_d_prev2 = (buffer_model.size() > 1) ? (float*)buffer_model[buffer_model.size() - 2]->data : vec_d_prev1;
+                        for (int j = 0; j < ggml_nelements(x_next); j++) {
+                            vec_x[j] += h_n * ((23 * vec_d_cur[j] - 16 * vec_d_prev1[j] + 5 * vec_d_prev2[j]) / 12);
+                        }
+                        break;
+                    }
+
+                    case 4: {
+                        float h_n_2 = (i > 1) ? (sigmas[i - 1] - sigmas[i - 2]) : h_n_1;
+                        float h_n_3 = (i > 2) ? (sigmas[i - 2] - sigmas[i - 3]) : h_n_2;
+                        float* vec_d_prev1 = (float*)buffer_model.back()->data;
+                        float* vec_d_prev2 = (buffer_model.size() > 1) ? (float*)buffer_model[buffer_model.size() - 2]->data : vec_d_prev1;
+                        float* vec_d_prev3 = (buffer_model.size() > 2) ? (float*)buffer_model[buffer_model.size() - 3]->data : vec_d_prev2;
+                        for (int j = 0; j < ggml_nelements(x_next); j++) {
+                            vec_x[j] += h_n * ((55 * vec_d_cur[j] - 59 * vec_d_prev1[j] + 37 * vec_d_prev2[j] - 9 * vec_d_prev3[j]) / 24);
+                        }
+                        break;
+                    }
+                }
+
+                // Manage buffer_model
+                if (buffer_model.size() == max_order - 1) {
+                    buffer_model.erase(buffer_model.begin());
+                }
+                buffer_model.push_back(d_cur);
+
+                // Prepare the next d tensor
+                d_cur = ggml_dup_tensor(work_ctx, x_next);
+            }
+        } break;
         case LCM:  // Latent Consistency Models
         {
             struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);

diffusion_model.hpp

@@ -1,9 +1,9 @@
 #ifndef __DIFFUSION_MODEL_H__
 #define __DIFFUSION_MODEL_H__
 
+#include "flux.hpp"
 #include "mmdit.hpp"
 #include "unet.hpp"
-#include "flux.hpp"
 
 struct DiffusionModel {
     virtual void compute(int n_threads,
@@ -124,13 +124,12 @@ struct MMDiTModel : public DiffusionModel {
     }
 };
 
-
 struct FluxModel : public DiffusionModel {
     Flux::FluxRunner flux;
 
     FluxModel(ggml_backend_t backend,
-               ggml_type wtype,
-               SDVersion version = VERSION_FLUX_DEV)
+              ggml_type wtype,
+              SDVersion version = VERSION_FLUX_DEV)
         : flux(backend, wtype, version) {
     }
 

docs/flux.md

@@ -4,14 +4,17 @@ You can run Flux using stable-diffusion.cpp with a GPU that has 6GB or even 4GB
 
 ## Download weights
 
-- Download flux-dev from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors
-- Download flux-schnell from https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/flux1-schnell.safetensors
+- Download flux
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf)
+    - Otherwise, download flux-dev from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors or flux-schnell from https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/flux1-schnell.safetensors
 - Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
 - Download clip_l from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
 - Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
 
 ## Convert flux weights
 
+You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf), this way you don't have to do the conversion yourself.
+
 Using fp16 will lead to overflow, but ggml's support for bf16 is not yet fully developed. Therefore, we need to convert flux to gguf format here, which also saves VRAM. For example:
 ```
 .\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
@@ -30,10 +33,10 @@ For example:
 
 Using formats of different precisions will yield results of varying quality.
 
-| Type | q8_0  | q4_0  | q3_k  | q2_k |
-|---- | ----  |----  |----  |----  |
-| **Memory** | 12068.09 MB  | 6394.53 MB  | 4888.16 MB  | 3735.73 MB |
-| **Result** | ![](../assets/flux/flux1-dev-q8_0.png) |![](../assets/flux/flux1-dev-q4_0.png) |![](../assets/flux/flux1-dev-q3_k.png) |![](../assets/flux/flux1-dev-q2_k.png)|
+| Type | q8_0  | q4_0  | q4_k  | q3_k  | q2_k |
+|---- | ----  |----  |----  |----  |----  |
+| **Memory** | 12068.09 MB  | 6394.53 MB | 6395.17 MB | 4888.16 MB  | 3735.73 MB |
+| **Result** | ![](../assets/flux/flux1-dev-q8_0.png) |![](../assets/flux/flux1-dev-q4_0.png) |![](../assets/flux/flux1-dev-q4_k.png) |![](../assets/flux/flux1-dev-q3_k.png) |![](../assets/flux/flux1-dev-q2_k.png)|
 
 
 

examples/cli/main.cpp

@@ -36,6 +36,8 @@ const char* sample_method_str[] = {
     "dpm++2s_a",
     "dpm++2m",
     "dpm++2mv2",
+    "ipndm",
+    "ipndm_v",
     "lcm",
 };
 
@@ -44,7 +46,9 @@ const char* schedule_str[] = {
     "default",
     "discrete",
     "karras",
+    "exponential",
     "ays",
+    "gits",
 };
 
 const char* modes_str[] = {
@@ -134,7 +138,7 @@ void print_params(SDParams params) {
     printf("    stacked_id_embeddings_path:   %s\n", params.stacked_id_embeddings_path.c_str());
     printf("    input_id_images_path:   %s\n", params.input_id_images_path.c_str());
     printf("    style ratio:       %.2f\n", params.style_ratio);
-    printf("    normzalize input image :  %s\n", params.normalize_input ? "true" : "false");
+    printf("    normalize input image :  %s\n", params.normalize_input ? "true" : "false");
     printf("    output_path:       %s\n", params.output_path.c_str());
     printf("    init_img:          %s\n", params.input_path.c_str());
     printf("    control_image:     %s\n", params.control_image_path.c_str());
@@ -169,7 +173,10 @@ void print_usage(int argc, const char* argv[]) {
     printf("  -M, --mode [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)\n");
     printf("  -t, --threads N                    number of threads to use during computation (default: -1).\n");
     printf("                                     If threads <= 0, then threads will be set to the number of CPU physical cores\n");
-    printf("  -m, --model [MODEL]                path to model\n");
+    printf("  -m, --model [MODEL]                path to full model\n");
+    printf("  --diffusion-model                  path to the standalone diffusion model\n");
+    printf("  --clip_l                           path to the clip-l text encoder\n");
+    printf("  --t5xxl                            path to the the t5xxl text encoder.\n");
     printf("  --vae [VAE]                        path to vae\n");
     printf("  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)\n");
     printf("  --control-net [CONTROL_PATH]       path to control net model\n");
@@ -194,16 +201,18 @@ void print_usage(int argc, const char* argv[]) {
     printf("                                     1.0 corresponds to full destruction of information in init image\n");
     printf("  -H, --height H                     image height, in pixel space (default: 512)\n");
     printf("  -W, --width W                      image width, in pixel space (default: 512)\n");
-    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, lcm}\n");
+    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm}\n");
     printf("                                     sampling method (default: \"euler_a\")\n");
     printf("  --steps  STEPS                     number of sample steps (default: 20)\n");
     printf("  --rng {std_default, cuda}          RNG (default: cuda)\n");
     printf("  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)\n");
     printf("  -b, --batch-count COUNT            number of images to generate.\n");
-    printf("  --schedule {discrete, karras, ays} Denoiser sigma schedule (default: discrete)\n");
+    printf("  --schedule {discrete, karras, exponential, ays, gits} Denoiser sigma schedule (default: discrete)\n");
     printf("  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)\n");
     printf("                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x\n");
     printf("  --vae-tiling                       process vae in tiles to reduce memory usage\n");
+    printf("  --vae-on-cpu                       keep vae in cpu (for low vram)\n");
+    printf("  --clip-on-cpu                      keep clip in cpu (for low vram).\n");
     printf("  --control-net-cpu                  keep controlnet in cpu (for low vram)\n");
     printf("  --canny                            apply canny preprocessor (edge detection)\n");
     printf("  --color                            Colors the logging tags according to level\n");

ggml_extend.hpp

@@ -32,6 +32,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef SD_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #ifdef SD_USE_SYCL
 #include "ggml-sycl.h"
 #endif
@@ -349,6 +353,12 @@ __STATIC_INLINE__ void ggml_split_tensor_2d(struct ggml_tensor* input,
     }
 }
 
+// unclamped -> expects x in the range [0-1]
+__STATIC_INLINE__ float ggml_smootherstep_f32(const float x) {
+    GGML_ASSERT(x >= 0.f && x <= 1.f);
+    return x * x * x * (x * (6.0f * x - 15.0f) + 10.0f);
+}
+
 __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
                                             struct ggml_tensor* output,
                                             int x,
@@ -357,6 +367,10 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
     int64_t width    = input->ne[0];
     int64_t height   = input->ne[1];
     int64_t channels = input->ne[2];
+
+    int64_t img_width    = output->ne[0];
+    int64_t img_height   = output->ne[1];
+
     GGML_ASSERT(input->type == GGML_TYPE_F32 && output->type == GGML_TYPE_F32);
     for (int iy = 0; iy < height; iy++) {
         for (int ix = 0; ix < width; ix++) {
@@ -364,16 +378,23 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
                 float new_value = ggml_tensor_get_f32(input, ix, iy, k);
                 if (overlap > 0) {  // blend colors in overlapped area
                     float old_value = ggml_tensor_get_f32(output, x + ix, y + iy, k);
-                    if (x > 0 && ix < overlap) {  // in overlapped horizontal
-                        ggml_tensor_set_f32(output, old_value + (new_value - old_value) * (ix / (1.0f * overlap)), x + ix, y + iy, k);
-                        continue;
-                    }
-                    if (y > 0 && iy < overlap) {  // in overlapped vertical
-                        ggml_tensor_set_f32(output, old_value + (new_value - old_value) * (iy / (1.0f * overlap)), x + ix, y + iy, k);
-                        continue;
-                    }
+
+                    const float x_f_0 = (x > 0) ? ix / float(overlap) : 1;
+                    const float x_f_1 = (x < (img_width - width)) ? (width - ix) / float(overlap) : 1 ;
+                    const float y_f_0 = (y > 0) ? iy / float(overlap) : 1;
+                    const float y_f_1 = (y < (img_height - height)) ? (height - iy) / float(overlap) : 1;
+
+                    const float x_f = std::min(std::min(x_f_0, x_f_1), 1.f);
+                    const float y_f = std::min(std::min(y_f_0, y_f_1), 1.f);
+
+                    ggml_tensor_set_f32(
+                        output,
+                        old_value + new_value * ggml_smootherstep_f32(y_f) * ggml_smootherstep_f32(x_f),
+                        x + ix, y + iy, k
+                    );
+                } else {
+                    ggml_tensor_set_f32(output, new_value, x + ix, y + iy, k);
                 }
-                ggml_tensor_set_f32(output, new_value, x + ix, y + iy, k);
             }
         }
     }
@@ -541,7 +562,7 @@ __STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const
 
 __STATIC_INLINE__ struct ggml_tensor* ggml_group_norm_32(struct ggml_context* ctx,
                                                          struct ggml_tensor* a) {
-    const float eps = 1e-6f; // default eps parameter
+    const float eps = 1e-6f;  // default eps parameter
     return ggml_group_norm(ctx, a, 32, eps);
 }
 
@@ -655,7 +676,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention(struct ggml_context* ctx
                                                         struct ggml_tensor* k,
                                                         struct ggml_tensor* v,
                                                         bool mask = false) {
-#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL) && !defined(SD_USE_SYCL)
+#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL) && !defined(SD_USE_VULKAN) && !defined(SD_USE_SYCL)
     struct ggml_tensor* kqv = ggml_flash_attn(ctx, q, k, v, false);  // [N * n_head, n_token, d_head]
 #else
     float d_head = (float)q->ne[0];
@@ -683,27 +704,27 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
                                                             int64_t n_head,
                                                             struct ggml_tensor* mask = NULL,
                                                             bool diag_mask_inf       = false,
-                                                            bool skip_reshape = false) {
+                                                            bool skip_reshape        = false) {
     int64_t L_q;
     int64_t L_k;
-    int64_t C  ;
-    int64_t N  ;
+    int64_t C;
+    int64_t N;
     int64_t d_head;
     if (!skip_reshape) {
-        L_q = q->ne[1];
-        L_k = k->ne[1];
-        C   = q->ne[0];
-        N   = q->ne[2];
+        L_q    = q->ne[1];
+        L_k    = k->ne[1];
+        C      = q->ne[0];
+        N      = q->ne[2];
         d_head = C / n_head;
-        q = ggml_reshape_4d(ctx, q, d_head, n_head, L_q, N);   // [N, L_q, n_head, d_head]
-        q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));  // [N, n_head, L_q, d_head]
-        q = ggml_reshape_3d(ctx, q, d_head, L_q, n_head * N);  // [N * n_head, L_q, d_head]
+        q      = ggml_reshape_4d(ctx, q, d_head, n_head, L_q, N);   // [N, L_q, n_head, d_head]
+        q      = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));  // [N, n_head, L_q, d_head]
+        q      = ggml_reshape_3d(ctx, q, d_head, L_q, n_head * N);  // [N * n_head, L_q, d_head]
 
         k = ggml_reshape_4d(ctx, k, d_head, n_head, L_k, N);   // [N, L_k, n_head, d_head]
         k = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3));  // [N, n_head, L_k, d_head]
         k = ggml_reshape_3d(ctx, k, d_head, L_k, n_head * N);  // [N * n_head, L_k, d_head]
 
-        v = ggml_reshape_4d(ctx, v, d_head, n_head, L_k, N);   // [N, L_k, n_head, d_head]
+        v = ggml_reshape_4d(ctx, v, d_head, n_head, L_k, N);  // [N, L_k, n_head, d_head]
     } else {
         L_q    = q->ne[1];
         L_k    = k->ne[1];
@@ -712,15 +733,15 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
         C      = d_head * n_head;
     }
 
-    float scale    = (1.0f / sqrt((float)d_head));
+    float scale = (1.0f / sqrt((float)d_head));
 
     bool use_flash_attn = false;
-    ggml_tensor* kqv = NULL;
+    ggml_tensor* kqv    = NULL;
     if (use_flash_attn) {
         v = ggml_cont(ctx, ggml_permute(ctx, v, 0, 2, 1, 3));  // [N, n_head, L_k, d_head]
         v = ggml_reshape_3d(ctx, v, d_head, L_k, n_head * N);  // [N * n_head, L_k, d_head]
         LOG_DEBUG("k->ne[1] == %d", k->ne[1]);
-        kqv = ggml_flash_attn_ext(ctx, q, k, v, mask, scale, 0);
+        kqv = ggml_flash_attn_ext(ctx, q, k, v, mask, scale, 0, 0);
     } else {
         v = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));  // [N, n_head, d_head, L_k]
         v = ggml_reshape_3d(ctx, v, L_k, d_head, n_head * N);  // [N * n_head, d_head, L_k]
@@ -770,8 +791,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_group_norm(struct ggml_context* ct
         b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
     }
 
-    const float eps = 1e-6f; // default eps parameter
-    x = ggml_group_norm(ctx, x, num_groups, eps);
+    const float eps = 1e-6f;  // default eps parameter
+    x               = ggml_group_norm(ctx, x, num_groups, eps);
     if (w != NULL && b != NULL) {
         x = ggml_mul(ctx, x, w);
         // b = ggml_repeat(ctx, b, x);
@@ -781,7 +802,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_group_norm(struct ggml_context* ct
 }
 
 __STATIC_INLINE__ void ggml_backend_tensor_get_and_sync(ggml_backend_t backend, const struct ggml_tensor* tensor, void* data, size_t offset, size_t size) {
-#if defined (SD_USE_CUBLAS) || defined (SD_USE_SYCL)
+#if defined(SD_USE_CUBLAS) || defined(SD_USE_SYCL)
     if (!ggml_backend_is_cpu(backend)) {
         ggml_backend_tensor_get_async(backend, tensor, data, offset, size);
         ggml_backend_synchronize(backend);
@@ -889,7 +910,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_timestep_embedding(
     struct ggml_context* ctx,
     struct ggml_tensor* timesteps,
     int dim,
-    int max_period = 10000,
+    int max_period    = 10000,
     float time_factor = 1.0f) {
     timesteps = ggml_scale(ctx, timesteps, time_factor);
     return ggml_timestep_embedding(ctx, timesteps, dim, max_period);
@@ -1187,9 +1208,10 @@ protected:
     int64_t in_features;
     int64_t out_features;
     bool bias;
+    bool force_f32;
 
     void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        if (in_features % ggml_blck_size(wtype) != 0) {
+        if (in_features % ggml_blck_size(wtype) != 0 || force_f32) {
             wtype = GGML_TYPE_F32;
         }
         params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
@@ -1201,10 +1223,12 @@ protected:
 public:
     Linear(int64_t in_features,
            int64_t out_features,
-           bool bias = true)
+           bool bias      = true,
+           bool force_f32 = false)
         : in_features(in_features),
           out_features(out_features),
-          bias(bias) {}
+          bias(bias),
+          force_f32(force_f32) {}
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
         struct ggml_tensor* w = params["weight"];

gits_noise.inl

@@ -0,0 +1,349 @@
+#ifndef GITS_NOISE_INL
+#define GITS_NOISE_INL
+
+const std::vector<std::vector<float>> GITS_NOISE_0_80 = {
+    { 14.61464119f, 7.49001646f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 6.77309084f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 3.07277966f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 2.05039096f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 2.05039096f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.19567990f, 1.98035145f, 0.86115354f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.19567990f, 1.98035145f, 0.86115354f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.88507891f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.07277966f, 1.84880662f, 0.83188516f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_0_85 = {
+    { 14.61464119f, 7.49001646f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 1.84880662f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 6.77309084f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.11996698f, 3.07277966f, 1.24153244f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 2.84484982f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.09240818f, 2.84484982f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.58536053f, 3.19567990f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 5.58536053f, 3.19567990f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.60512662f, 2.63833880f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.88507891f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_0_90 = {
+    { 14.61464119f, 6.77309084f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 3.07277966f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.54230714f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.54230714f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 3.07277966f, 1.61558151f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.11996698f, 4.86714602f, 3.07277966f, 1.61558151f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 2.95596409f, 1.61558151f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.84484982f, 1.84880662f, 1.08895338f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.84484982f, 1.84880662f, 1.08895338f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.45427561f, 3.32507086f, 2.45070267f, 1.61558151f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.45427561f, 3.32507086f, 2.45070267f, 1.61558151f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.19988537f, 1.51179266f, 0.89115214f, 0.43325692f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_0_95 = {
+    { 14.61464119f, 6.77309084f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 2.84484982f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.36326075f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.56271636f, 0.64427125f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.95596409f, 1.56271636f, 0.64427125f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.41535246f, 0.803307f, 0.38853383f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.46139455f, 2.63833880f, 1.84880662f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.46139455f, 2.63833880f, 1.84880662f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.60512662f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.60512662f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.75677586f, 3.07277966f, 2.45070267f, 1.78698075f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.36326075f, 1.72759056f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.36326075f, 1.72759056f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.75677586f, 3.07277966f, 2.45070267f, 1.91321158f, 1.46270394f, 1.05362725f, 0.72133851f, 0.43325692f, 0.19894916f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_00 = {
+    { 14.61464119f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.36326075f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 7.11996698f, 3.07277966f, 1.56271636f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.41535246f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.86115354f, 0.38853383f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.86115354f, 0.38853383f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 3.07277966f, 1.98035145f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.98035145f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.27973175f, 1.51179266f, 0.95350921f, 0.54755926f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.12350607f, 1.56271636f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.61558151f, 1.162866f, 0.803307f, 0.50118381f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.75677586f, 3.07277966f, 2.45070267f, 1.84880662f, 1.36964464f, 1.01931262f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.46139455f, 2.84484982f, 2.19988537f, 1.67050016f, 1.24153244f, 0.92192322f, 0.64427125f, 0.43325692f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_05 = {
+    { 14.61464119f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.05039096f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.84484982f, 1.28281462f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.61558151f, 0.803307f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.56271636f, 0.803307f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.95350921f, 0.52423614f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 1.98035145f, 1.24153244f, 0.74807048f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.27973175f, 1.51179266f, 0.95350921f, 0.59516323f, 0.34370604f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.09240818f, 3.46139455f, 2.45070267f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 3.46139455f, 2.45070267f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.72759056f, 1.24153244f, 0.86115354f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.61558151f, 1.162866f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.67050016f, 1.28281462f, 0.95350921f, 0.72133851f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.36326075f, 1.84880662f, 1.41535246f, 1.08895338f, 0.83188516f, 0.61951244f, 0.45573691f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.57119018f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.30717278f, 7.11996698f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.57119018f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.30717278f, 7.11996698f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.98035145f, 1.61558151f, 1.32549286f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.41087446f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_10 = {
+    { 14.61464119f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 1.61558151f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.45070267f, 1.08895338f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.95596409f, 1.56271636f, 0.803307f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.61558151f, 0.89115214f, 0.4783645f, 0.19894916f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.84880662f, 1.08895338f, 0.64427125f, 0.34370604f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.95350921f, 0.54755926f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.91321158f, 1.24153244f, 0.803307f, 0.4783645f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.05039096f, 1.41535246f, 0.95350921f, 0.64427125f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.27973175f, 1.61558151f, 1.12534678f, 0.803307f, 0.54755926f, 0.36617002f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.32507086f, 2.45070267f, 1.72759056f, 1.24153244f, 0.89115214f, 0.64427125f, 0.45573691f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.09240818f, 3.60512662f, 2.84484982f, 2.05039096f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.09240818f, 3.60512662f, 2.84484982f, 2.12350607f, 1.61558151f, 1.24153244f, 0.95350921f, 0.72133851f, 0.54755926f, 0.41087446f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.43325692f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.43325692f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.89115214f, 0.72133851f, 0.59516323f, 0.4783645f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_15 = {
+    { 14.61464119f, 0.83188516f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 1.56271636f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 1.91321158f, 0.83188516f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.24153244f, 0.59516323f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.51179266f, 0.803307f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.56271636f, 0.89115214f, 0.50118381f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.84880662f, 1.12534678f, 0.72133851f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.91321158f, 1.24153244f, 0.803307f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.91321158f, 1.24153244f, 0.803307f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.05039096f, 1.36964464f, 0.95350921f, 0.69515091f, 0.4783645f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.78698075f, 1.32549286f, 1.01931262f, 0.803307f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.78698075f, 1.32549286f, 1.01931262f, 0.803307f, 0.64427125f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.12534678f, 0.89115214f, 0.72133851f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.12534678f, 0.89115214f, 0.72133851f, 0.59516323f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_20 = {
+    { 14.61464119f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 0.92192322f, 0.36617002f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.24153244f, 0.59516323f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.05039096f, 0.95350921f, 0.45573691f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.24153244f, 0.64427125f, 0.29807833f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.803307f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 0.95350921f, 0.59516323f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.67050016f, 1.08895338f, 0.74807048f, 0.50118381f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.24153244f, 0.83188516f, 0.59516323f, 0.41087446f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 1.98035145f, 1.36964464f, 0.95350921f, 0.69515091f, 0.50118381f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.46139455f, 2.36326075f, 1.56271636f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.46139455f, 2.45070267f, 1.61558151f, 1.162866f, 0.86115354f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.20157266f, 0.92192322f, 0.72133851f, 0.57119018f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_25 = {
+    { 14.61464119f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 2.05039096f, 0.803307f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 0.95350921f, 0.43325692f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.24153244f, 0.59516323f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.51179266f, 0.803307f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.36326075f, 1.24153244f, 0.72133851f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.83188516f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 0.98595673f, 0.64427125f, 0.43325692f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.67050016f, 1.08895338f, 0.74807048f, 0.52423614f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.24153244f, 0.86115354f, 0.64427125f, 0.4783645f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.28281462f, 0.92192322f, 0.69515091f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.72133851f, 0.54755926f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.72133851f, 0.57119018f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.74807048f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.41535246f, 1.05362725f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.41535246f, 1.05362725f, 0.803307f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.46270394f, 1.08895338f, 0.83188516f, 0.66947293f, 0.54755926f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_30 = {
+    { 14.61464119f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 1.24153244f, 0.43325692f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.59516323f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.803307f, 0.36617002f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.01931262f, 0.52423614f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.36964464f, 0.74807048f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.56271636f, 0.89115214f, 0.54755926f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 0.95350921f, 0.61951244f, 0.41087446f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.83188516f, 0.54755926f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.41535246f, 0.92192322f, 0.64427125f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.56271636f, 1.01931262f, 0.72133851f, 0.50118381f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.05362725f, 0.74807048f, 0.54755926f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.77538133f, 0.57119018f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.83188516f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.78698075f, 1.24153244f, 0.92192322f, 0.72133851f, 0.57119018f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.78698075f, 1.24153244f, 0.92192322f, 0.72133851f, 0.57119018f, 0.4783645f, 0.41087446f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_35 = {
+    { 14.61464119f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 0.95350921f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.57119018f, 0.19894916f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.69515091f, 0.29807833f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.83188516f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.162866f, 0.64427125f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.36964464f, 0.803307f, 0.50118381f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.41535246f, 0.83188516f, 0.54755926f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.95350921f, 0.64427125f, 0.45573691f, 0.32104823f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.95350921f, 0.64427125f, 0.45573691f, 0.34370604f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 1.01931262f, 0.72133851f, 0.52423614f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 1.01931262f, 0.72133851f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 1.05362725f, 0.74807048f, 0.54755926f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.72759056f, 1.12534678f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.72759056f, 1.12534678f, 0.803307f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.51179266f, 1.01931262f, 0.74807048f, 0.57119018f, 0.45573691f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_40 = {
+    { 14.61464119f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 0.95350921f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 1.08895338f, 0.43325692f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.64427125f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.803307f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.05039096f, 0.95350921f, 0.54755926f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.72133851f, 0.43325692f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.52423614f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.54755926f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.41535246f, 0.86115354f, 0.59516323f, 0.43325692f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.64427125f, 0.45573691f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.64427125f, 0.4783645f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.98595673f, 0.69515091f, 0.52423614f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.72133851f, 0.54755926f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.05362725f, 0.74807048f, 0.57119018f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.43325692f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.45573691f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_45 = {
+    { 14.61464119f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 0.803307f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 0.95350921f, 0.34370604f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.24153244f, 0.54755926f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.72133851f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.803307f, 0.45573691f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.91321158f, 0.95350921f, 0.57119018f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.19988537f, 1.08895338f, 0.64427125f, 0.41087446f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.34370604f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.36617002f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.54755926f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.83188516f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.83188516f, 0.59516323f, 0.45573691f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.69515091f, 0.52423614f, 0.41087446f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.69515091f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.98595673f, 0.72133851f, 0.54755926f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.74807048f, 0.57119018f, 0.4783645f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.74807048f, 0.59516323f, 0.50118381f, 0.43325692f, 0.38853383f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_50 = {
+    { 14.61464119f, 0.54755926f, 0.02916753f },
+    { 14.61464119f, 0.803307f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 0.86115354f, 0.32104823f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.24153244f, 0.54755926f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.72133851f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.803307f, 0.45573691f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.83188516f, 0.52423614f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.95350921f, 0.59516323f, 0.38853383f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.95350921f, 0.59516323f, 0.41087446f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.95350921f, 0.61951244f, 0.43325692f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.19988537f, 1.12534678f, 0.72133851f, 0.50118381f, 0.36617002f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.19988537f, 1.12534678f, 0.72133851f, 0.50118381f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.59516323f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.32549286f, 0.86115354f, 0.64427125f, 0.50118381f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.36964464f, 0.92192322f, 0.69515091f, 0.54755926f, 0.45573691f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.41535246f, 0.95350921f, 0.72133851f, 0.57119018f, 0.4783645f, 0.43325692f, 0.38853383f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<const std::vector<std::vector<float>>*> GITS_NOISE = {
+    { &GITS_NOISE_0_80 },
+    { &GITS_NOISE_0_85 },
+    { &GITS_NOISE_0_90 },
+    { &GITS_NOISE_0_95 },
+    { &GITS_NOISE_1_00 },
+    { &GITS_NOISE_1_05 },
+    { &GITS_NOISE_1_10 },
+    { &GITS_NOISE_1_15 },
+    { &GITS_NOISE_1_20 },
+    { &GITS_NOISE_1_25 },
+    { &GITS_NOISE_1_30 },
+    { &GITS_NOISE_1_35 },
+    { &GITS_NOISE_1_40 },
+    { &GITS_NOISE_1_45 },
+    { &GITS_NOISE_1_50 }
+};
+
+#endif // GITS_NOISE_INL

lora.hpp

@@ -10,10 +10,10 @@ struct LoraModel : public GGMLRunner {
     std::map<std::string, struct ggml_tensor*> lora_tensors;
     std::string file_path;
     ModelLoader model_loader;
-    bool load_failed = false;
-    bool applied     = false;
+    bool load_failed                = false;
+    bool applied                    = false;
     std::vector<int> zero_index_vec = {0};
-    ggml_tensor* zero_index = NULL;
+    ggml_tensor* zero_index         = NULL;
 
     LoraModel(ggml_backend_t backend,
               ggml_type wtype,
@@ -72,8 +72,8 @@ struct LoraModel : public GGMLRunner {
 
     ggml_tensor* to_f32(ggml_context* ctx, ggml_tensor* a) {
         auto out = ggml_reshape_1d(ctx, a, ggml_nelements(a));
-        out = ggml_get_rows(ctx, out, zero_index);
-        out = ggml_reshape(ctx, out, a);
+        out      = ggml_get_rows(ctx, out, zero_index);
+        out      = ggml_reshape(ctx, out, a);
         return out;
     }
 
@@ -82,6 +82,7 @@ struct LoraModel : public GGMLRunner {
 
         zero_index = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, 1);
         set_backend_tensor_data(zero_index, zero_index_vec.data());
+        ggml_build_forward_expand(gf, zero_index);
 
         std::set<std::string> applied_lora_tensors;
         for (auto it : model_tensors) {

mmdit.hpp

@@ -101,8 +101,8 @@ public:
     TimestepEmbedder(int64_t hidden_size,
                      int64_t frequency_embedding_size = 256)
         : frequency_embedding_size(frequency_embedding_size) {
-        blocks["mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(frequency_embedding_size, hidden_size));
-        blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size));
+        blocks["mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(frequency_embedding_size, hidden_size, true, true));
+        blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size, true, true));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* t) {
@@ -125,8 +125,8 @@ struct VectorEmbedder : public GGMLBlock {
 public:
     VectorEmbedder(int64_t input_dim,
                    int64_t hidden_size) {
-        blocks["mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(input_dim, hidden_size));
-        blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size));
+        blocks["mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(input_dim, hidden_size, true, true));
+        blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size, true, true));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
@@ -423,7 +423,7 @@ public:
                int64_t out_channels) {
         // total_out_channels is always None
         blocks["norm_final"]         = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
-        blocks["linear"]             = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, patch_size * patch_size * out_channels));
+        blocks["linear"]             = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, patch_size * patch_size * out_channels, true, true));
         blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size));
     }
 
@@ -510,7 +510,7 @@ public:
             blocks["y_embedder"] = std::shared_ptr<GGMLBlock>(new VectorEmbedder(adm_in_channels, hidden_size));
         }
 
-        blocks["context_embedder"] = std::shared_ptr<GGMLBlock>(new Linear(4096, 1536));
+        blocks["context_embedder"] = std::shared_ptr<GGMLBlock>(new Linear(4096, 1536, true, true));
 
         for (int i = 0; i < depth; i++) {
             blocks["joint_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new JointBlock(hidden_size,

model.cpp

@@ -21,6 +21,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef SD_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #define ST_HEADER_SIZE_LEN 8
 
 uint64_t read_u64(uint8_t* buffer) {
@@ -567,10 +571,10 @@ uint16_t f8_e4m3_to_f16(uint8_t f8) {
         return ggml_fp32_to_fp16(NAN);
     }
 
-    uint32_t sign = f8 & 0x80;
+    uint32_t sign     = f8 & 0x80;
     uint32_t exponent = (f8 & 0x78) >> 3;
     uint32_t mantissa = f8 & 0x07;
-    uint32_t result = sign << 24;
+    uint32_t result   = sign << 24;
     if (exponent == 0) {
         if (mantissa > 0) {
             exponent = 0x7f - exponent_bias;
@@ -1397,10 +1401,11 @@ ggml_type ModelLoader::get_sd_wtype() {
             continue;
         }
 
-        if (tensor_storage.name.find(".weight") != std::string::npos &&
-            (tensor_storage.name.find("time_embed") != std::string::npos ||
-            tensor_storage.name.find("context_embedder") != std::string::npos ||
-            tensor_storage.name.find("time_in") != std::string::npos)) {
+        if (ggml_is_quantized(tensor_storage.type)) {
+            return tensor_storage.type;
+        }
+
+        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
             return tensor_storage.type;
         }
     }
@@ -1414,20 +1419,23 @@ ggml_type ModelLoader::get_conditioner_wtype() {
         }
 
         if ((tensor_storage.name.find("text_encoders") == std::string::npos &&
-            tensor_storage.name.find("cond_stage_model") == std::string::npos &&
-            tensor_storage.name.find("te.text_model.") == std::string::npos &&
-            tensor_storage.name.find("conditioner") == std::string::npos)) {
+             tensor_storage.name.find("cond_stage_model") == std::string::npos &&
+             tensor_storage.name.find("te.text_model.") == std::string::npos &&
+             tensor_storage.name.find("conditioner") == std::string::npos)) {
             continue;
         }
 
-        if (tensor_storage.name.find(".weight") != std::string::npos) {
+        if (ggml_is_quantized(tensor_storage.type)) {
+            return tensor_storage.type;
+        }
+
+        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
             return tensor_storage.type;
         }
     }
     return GGML_TYPE_COUNT;
 }
 
-
 ggml_type ModelLoader::get_diffusion_model_wtype() {
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
@@ -1438,10 +1446,11 @@ ggml_type ModelLoader::get_diffusion_model_wtype() {
             continue;
         }
 
-        if (tensor_storage.name.find(".weight") != std::string::npos &&
-            (tensor_storage.name.find("time_embed") != std::string::npos ||
-            tensor_storage.name.find("context_embedder") != std::string::npos ||
-            tensor_storage.name.find("time_in") != std::string::npos)) {
+        if (ggml_is_quantized(tensor_storage.type)) {
+            return tensor_storage.type;
+        }
+
+        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
             return tensor_storage.type;
         }
     }
@@ -1459,7 +1468,11 @@ ggml_type ModelLoader::get_vae_wtype() {
             continue;
         }
 
-        if (tensor_storage.name.find(".weight")) {
+        if (ggml_is_quantized(tensor_storage.type)) {
+            return tensor_storage.type;
+        }
+
+        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
             return tensor_storage.type;
         }
     }
@@ -1724,6 +1737,37 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
     return true;
 }
 
+bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type) {
+    const std::string& name = tensor_storage.name;
+    if (type != GGML_TYPE_COUNT) {
+        if (ggml_is_quantized(type) && tensor_storage.ne[0] % ggml_blck_size(type) != 0) {
+            // Pass, do not convert
+        } else if (ends_with(name, ".bias")) {
+            // Pass, do not convert
+        } else if (ends_with(name, ".scale")) {
+            // Pass, do not convert
+        } else if (contains(name, "img_in.") ||
+                   contains(name, "txt_in.") ||
+                   contains(name, "time_in.") ||
+                   contains(name, "vector_in.") ||
+                   contains(name, "guidance_in.") ||
+                   contains(name, "final_layer.")) {
+            // Pass, do not convert. For FLUX
+        } else if (contains(name, "x_embedder.") ||
+                   contains(name, "t_embedder.") ||
+                   contains(name, "y_embedder.") ||
+                   contains(name, "pos_embed") ||
+                   contains(name, "context_embedder.")) {
+            // Pass, do not convert. For MMDiT
+        } else if (contains(name, "time_embed.") || contains(name, "label_emb.")) {
+            // Pass, do not convert. For Unet
+        } else {
+            return true;
+        }
+    }
+    return false;
+}
+
 bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type) {
     auto backend    = ggml_backend_cpu_init();
     size_t mem_size = 1 * 1024 * 1024;  // for padding
@@ -1738,12 +1782,8 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
         const std::string& name = tensor_storage.name;
 
         ggml_type tensor_type = tensor_storage.type;
-        if (type != GGML_TYPE_COUNT) {
-            if (ggml_is_quantized(type) && tensor_storage.ne[0] % ggml_blck_size(type) != 0) {
-                tensor_type = GGML_TYPE_F16;
-            } else {
-                tensor_type = type;
-            }
+        if (tensor_should_be_converted(tensor_storage, type)) {
+            tensor_type = type;
         }
 
         ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
@@ -1793,15 +1833,9 @@ int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type)
     }
 
     for (auto& tensor_storage : processed_tensor_storages) {
-        ggml_type tensor_type = tensor_storage.type;
-        if (type != GGML_TYPE_COUNT) {
-            if (ggml_is_quantized(type) && tensor_storage.ne[0] % 32 != 0) {
-                tensor_type = GGML_TYPE_F16;
-            } else {
-                tensor_type = type;
-            }
+        if (tensor_should_be_converted(tensor_storage, type)) {
+            tensor_storage.type = type;
         }
-        tensor_storage.type = tensor_type;
         mem_size += tensor_storage.nbytes() + alignment;
     }
 

model.h

@@ -157,6 +157,7 @@ public:
                       ggml_backend_t backend,
                       std::set<std::string> ignore_tensors = {});
     bool save_to_gguf_file(const std::string& file_path, ggml_type type);
+    bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;
 
@@ -165,4 +166,3 @@ public:
 };
 
 #endif  // __MODEL_H__
-

stable-diffusion.cpp

@@ -41,6 +41,8 @@ const char* sampling_methods_str[] = {
     "DPM++ (2s)",
     "DPM++ (2M)",
     "modified DPM++ (2M)",
+    "iPNDM",
+    "iPNDM_v",
     "LCM",
 };
 
@@ -69,11 +71,10 @@ public:
     ggml_backend_t clip_backend        = NULL;
     ggml_backend_t control_net_backend = NULL;
     ggml_backend_t vae_backend         = NULL;
-    ggml_type model_wtype          = GGML_TYPE_COUNT;
-    ggml_type conditioner_wtype          = GGML_TYPE_COUNT;
-    ggml_type diffusion_model_wtype          = GGML_TYPE_COUNT;
-    ggml_type vae_wtype          = GGML_TYPE_COUNT;
-
+    ggml_type model_wtype              = GGML_TYPE_COUNT;
+    ggml_type conditioner_wtype        = GGML_TYPE_COUNT;
+    ggml_type diffusion_model_wtype    = GGML_TYPE_COUNT;
+    ggml_type vae_wtype                = GGML_TYPE_COUNT;
 
     SDVersion version;
     bool vae_decode_only         = false;
@@ -161,6 +162,15 @@ public:
         ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
+#ifdef SD_USE_VULKAN
+        LOG_DEBUG("Using Vulkan backend");
+        for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
+            backend = ggml_backend_vk_init(device);
+        }
+        if(!backend) {
+            LOG_WARN("Failed to initialize Vulkan backend");
+        }
+#endif
 #ifdef SD_USE_SYCL
         LOG_DEBUG("Using SYCL backend");
         backend = ggml_backend_sycl_init(0);
@@ -171,7 +181,7 @@ public:
             backend = ggml_backend_cpu_init();
         }
 #ifdef SD_USE_FLASH_ATTENTION
-#if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL) || defined (SD_USE_SYCL)
+#if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL) || defined (SD_USE_SYCL) || defined(SD_USE_VULKAN)
         LOG_WARN("Flash Attention not supported with GPU Backend");
 #else
         LOG_INFO("Flash Attention enabled");
@@ -243,10 +253,10 @@ public:
                 vae_wtype = wtype;
             }
         } else {
-            model_wtype = wtype;
-            conditioner_wtype = wtype;
+            model_wtype           = wtype;
+            conditioner_wtype     = wtype;
             diffusion_model_wtype = wtype;
-            vae_wtype = wtype;
+            vae_wtype             = wtype;
         }
 
         if (version == VERSION_SDXL) {
@@ -255,7 +265,7 @@ public:
 
         LOG_INFO("Weight type:                 %s", ggml_type_name(model_wtype));
         LOG_INFO("Conditioner weight type:     %s", ggml_type_name(conditioner_wtype));
-        LOG_INFO("Diffsuion model weight type: %s", ggml_type_name(diffusion_model_wtype));
+        LOG_INFO("Diffusion model weight type: %s", ggml_type_name(diffusion_model_wtype));
         LOG_INFO("VAE weight type:             %s", ggml_type_name(vae_wtype));
 
         LOG_DEBUG("ggml tensor size = %d bytes", (int)sizeof(ggml_tensor));
@@ -290,7 +300,7 @@ public:
             first_stage_model->alloc_params_buffer();
             first_stage_model->get_param_tensors(tensors, "first_stage_model");
         } else {
-            clip_backend = backend;
+            clip_backend   = backend;
             bool use_t5xxl = false;
             if (version == VERSION_SD3_2B || version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
                 use_t5xxl = true;
@@ -508,7 +518,7 @@ public:
             LOG_INFO("running in Flux FLOW mode");
             float shift = 1.15f;
             if (version == VERSION_FLUX_SCHNELL) {
-                shift = 1.0f; // TODO: validate
+                shift = 1.0f;  // TODO: validate
             }
             denoiser = std::make_shared<FluxFlowDenoiser>(shift);
         } else if (is_using_v_parameterization) {
@@ -528,11 +538,20 @@ public:
                     LOG_INFO("running with Karras schedule");
                     denoiser->schedule = std::make_shared<KarrasSchedule>();
                     break;
+                case EXPONENTIAL:
+                    LOG_INFO("running exponential schedule");
+                    denoiser->schedule = std::make_shared<ExponentialSchedule>();
+                    break;
                 case AYS:
                     LOG_INFO("Running with Align-Your-Steps schedule");
                     denoiser->schedule          = std::make_shared<AYSSchedule>();
                     denoiser->schedule->version = version;
                     break;
+                case GITS:
+                    LOG_INFO("Running with GITS schedule");
+                    denoiser->schedule          = std::make_shared<GITSSchedule>();
+                    denoiser->schedule->version = version;
+                    break;
                 case DEFAULT:
                     // Don't touch anything.
                     break;

stable-diffusion.h

@@ -41,6 +41,8 @@ enum sample_method_t {
     DPMPP2S_A,
     DPMPP2M,
     DPMPP2Mv2,
+    IPNDM,
+    IPNDM_V,
     LCM,
     N_SAMPLE_METHODS
 };
@@ -49,43 +51,45 @@ enum schedule_t {
     DEFAULT,
     DISCRETE,
     KARRAS,
+    EXPONENTIAL,
     AYS,
+    GITS,
     N_SCHEDULES
 };
 
 // same as enum ggml_type
 enum sd_type_t {
-    SD_TYPE_F32     = 0,
-    SD_TYPE_F16     = 1,
-    SD_TYPE_Q4_0    = 2,
-    SD_TYPE_Q4_1    = 3,
+    SD_TYPE_F32  = 0,
+    SD_TYPE_F16  = 1,
+    SD_TYPE_Q4_0 = 2,
+    SD_TYPE_Q4_1 = 3,
     // SD_TYPE_Q4_2 = 4, support has been removed
     // SD_TYPE_Q4_3 = 5, support has been removed
-    SD_TYPE_Q5_0    = 6,
-    SD_TYPE_Q5_1    = 7,
-    SD_TYPE_Q8_0    = 8,
-    SD_TYPE_Q8_1    = 9,
-    SD_TYPE_Q2_K    = 10,
-    SD_TYPE_Q3_K    = 11,
-    SD_TYPE_Q4_K    = 12,
-    SD_TYPE_Q5_K    = 13,
-    SD_TYPE_Q6_K    = 14,
-    SD_TYPE_Q8_K    = 15,
-    SD_TYPE_IQ2_XXS = 16,
-    SD_TYPE_IQ2_XS  = 17,
-    SD_TYPE_IQ3_XXS = 18,
-    SD_TYPE_IQ1_S   = 19,
-    SD_TYPE_IQ4_NL  = 20,
-    SD_TYPE_IQ3_S   = 21,
-    SD_TYPE_IQ2_S   = 22,
-    SD_TYPE_IQ4_XS  = 23,
-    SD_TYPE_I8      = 24,
-    SD_TYPE_I16     = 25,
-    SD_TYPE_I32     = 26,
-    SD_TYPE_I64     = 27,
-    SD_TYPE_F64     = 28,
-    SD_TYPE_IQ1_M   = 29,
-    SD_TYPE_BF16    = 30,
+    SD_TYPE_Q5_0     = 6,
+    SD_TYPE_Q5_1     = 7,
+    SD_TYPE_Q8_0     = 8,
+    SD_TYPE_Q8_1     = 9,
+    SD_TYPE_Q2_K     = 10,
+    SD_TYPE_Q3_K     = 11,
+    SD_TYPE_Q4_K     = 12,
+    SD_TYPE_Q5_K     = 13,
+    SD_TYPE_Q6_K     = 14,
+    SD_TYPE_Q8_K     = 15,
+    SD_TYPE_IQ2_XXS  = 16,
+    SD_TYPE_IQ2_XS   = 17,
+    SD_TYPE_IQ3_XXS  = 18,
+    SD_TYPE_IQ1_S    = 19,
+    SD_TYPE_IQ4_NL   = 20,
+    SD_TYPE_IQ3_S    = 21,
+    SD_TYPE_IQ2_S    = 22,
+    SD_TYPE_IQ4_XS   = 23,
+    SD_TYPE_I8       = 24,
+    SD_TYPE_I16      = 25,
+    SD_TYPE_I32      = 26,
+    SD_TYPE_I64      = 27,
+    SD_TYPE_F64      = 28,
+    SD_TYPE_IQ1_M    = 29,
+    SD_TYPE_BF16     = 30,
     SD_TYPE_Q4_0_4_4 = 31,
     SD_TYPE_Q4_0_4_8 = 32,
     SD_TYPE_Q4_0_8_8 = 33,
@@ -203,7 +207,7 @@ SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
 
 SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor);
 
-SD_API bool convert(const char* input_path, const char* vae_path, const char* output_path, sd_type_t output_type);
+SD_API bool convert(const char* input_path, const char* vae_path, const char* output_path, enum sd_type_t output_type);
 
 SD_API uint8_t* preprocess_canny(uint8_t* img,
                                  int width,

upscaler.cpp

@@ -24,6 +24,10 @@ struct UpscalerGGML {
         ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
+#ifdef SD_USE_VULKAN
+        LOG_DEBUG("Using Vulkan backend");
+        backend = ggml_backend_vk_init(0);
+#endif
 #ifdef SD_USE_SYCL
         LOG_DEBUG("Using SYCL backend");
         backend = ggml_backend_sycl_init(0);