Merge branch 'leejet:master' into patch-2

Author: vmobilis

CMakeLists.txt

@@ -96,6 +96,7 @@ endif()
 if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
     add_definitions(-DSD_USE_SYCL)
     # disable fast-math on host, see:
     # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html

Dockerfile.musa

@@ -1,4 +1,4 @@
-ARG MUSA_VERSION=rc3.1.0
+ARG MUSA_VERSION=rc3.1.1
 
 FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu22.04 as build
 

README.md

@@ -326,7 +326,7 @@ These projects use `stable-diffusion.cpp` as a backend for their image generatio
 
 - [Jellybox](https://jellybox.com)
 - [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
-- [Stable Diffusion CLI-GUI] (https://github.com/piallai/stable-diffusion.cpp)
+- [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
 
 ## Contributors
 

denoiser.hpp

@@ -39,6 +39,8 @@ const char* sample_method_str[] = {
     "ipndm",
     "ipndm_v",
     "lcm",
+    "ddim_trailing",
+    "tcd",
 };
 
 // Names of the sigma schedule overrides, same order as sample_schedule in stable-diffusion.h
@@ -93,6 +95,7 @@ struct SDParams {
     float min_cfg     = 1.0f;
     float cfg_scale   = 7.0f;
     float guidance    = 3.5f;
+    float eta         = 0.f;
     float style_ratio = 20.f;
     int clip_skip     = -1;  // <= 0 represents unspecified
     int width         = 512;
@@ -162,6 +165,7 @@ void print_params(SDParams params) {
     printf("    cfg_scale:         %.2f\n", params.cfg_scale);
     printf("    slg_scale:         %.2f\n", params.slg_scale);
     printf("    guidance:          %.2f\n", params.guidance);
+    printf("    eta:               %.2f\n", params.eta);
     printf("    clip_skip:         %d\n", params.clip_skip);
     printf("    width:             %d\n", params.width);
     printf("    height:            %d\n", params.height);
@@ -202,13 +206,16 @@ void print_usage(int argc, const char* argv[]) {
     printf("                                     If not specified, the default is the type of the weight file\n");
     printf("  --lora-model-dir [DIR]             lora model directory\n");
     printf("  -i, --init-img [IMAGE]             path to the input image, required by img2img\n");
+    printf("  --mask [MASK]                      path to the mask image, required by img2img with mask\n");
     printf("  --control-image [IMAGE]            path to image condition, control net\n");
     printf("  -o, --output OUTPUT                path to write result image to (default: ./output.png)\n");
     printf("  -p, --prompt [PROMPT]              the prompt to render\n");
     printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
     printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
+    printf("  --guidance SCALE                   guidance scale for img2img (default: 3.5)\n");
     printf("  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)\n");
     printf("                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium\n");
+    printf("  --eta SCALE                        eta in DDIM, only for DDIM and TCD: (default: 0)\n");
     printf("  --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])\n");
     printf("  --skip-layer-start START           SLG enabling point: (default: 0.01)\n");
     printf("  --skip-layer-end END               SLG disabling point: (default: 0.2)\n");
@@ -219,7 +226,7 @@ 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, ipndm, ipndm_v, lcm}\n");
+    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}\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");
@@ -438,6 +445,12 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.guidance = std::stof(argv[i]);
+        } else if (arg == "--eta") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.eta = std::stof(argv[i]);
         } else if (arg == "--strength") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -717,6 +730,7 @@ std::string get_image_params(SDParams params, int64_t seed) {
         parameter_string += "Skip layer end: " + std::to_string(params.skip_layer_end) + ", ";
     }
     parameter_string += "Guidance: " + std::to_string(params.guidance) + ", ";
+    parameter_string += "Eta: " + std::to_string(params.eta) + ", ";
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(params.width) + "x" + std::to_string(params.height) + ", ";
     parameter_string += "Model: " + sd_basename(params.model_path) + ", ";
@@ -937,6 +951,7 @@ int main(int argc, const char* argv[]) {
                           params.clip_skip,
                           params.cfg_scale,
                           params.guidance,
+                          params.eta,
                           params.width,
                           params.height,
                           params.sample_method,
@@ -1004,6 +1019,7 @@ int main(int argc, const char* argv[]) {
                               params.clip_skip,
                               params.cfg_scale,
                               params.guidance,
+                              params.eta,
                               params.width,
                               params.height,
                               params.sample_method,

examples/cli/main.cpp

@@ -52,6 +52,71 @@
 #define __STATIC_INLINE__ static inline
 #endif
 
+// n-mode trensor-matrix product
+// example: 2-mode product
+// A: [ne03, k, ne01, ne00]
+// B: k rows, m columns => [k, m]
+// result is [ne03, m, ne01, ne00]
+__STATIC_INLINE__ struct ggml_tensor* ggml_mul_n_mode(struct ggml_context* ctx, struct ggml_tensor* a, struct ggml_tensor* b, int mode = 0) {
+    // reshape A
+    // swap 0th and nth axis
+    a       = ggml_cont(ctx, ggml_permute(ctx, a, mode, mode != 1 ? 1 : 0, mode != 2 ? 2 : 0, mode != 3 ? 3 : 0));
+    int ne1 = a->ne[1];
+    int ne2 = a->ne[2];
+    int ne3 = a->ne[3];
+    // make 2D
+    a = ggml_cont(ctx, ggml_reshape_2d(ctx, a, a->ne[0], (ne3 * ne2 * ne1)));
+
+    struct ggml_tensor* result = ggml_cont(ctx, ggml_transpose(ctx, ggml_mul_mat(ctx, a, b)));
+
+    // reshape output (same shape as a after permutation except first dim)
+    result = ggml_reshape_4d(ctx, result, result->ne[0], ne1, ne2, ne3);
+    // swap back 0th and nth axis
+    result = ggml_permute(ctx, result, mode, mode != 1 ? 1 : 0, mode != 2 ? 2 : 0, mode != 3 ? 3 : 0);
+    return result;
+}
+
+__STATIC_INLINE__ struct ggml_tensor* ggml_merge_lora(ggml_context* ctx, struct ggml_tensor* lora_down, struct ggml_tensor* lora_up, struct ggml_tensor* lora_mid = NULL) {
+    struct ggml_tensor* updown;
+    // flat lora tensors to multiply it
+    int64_t lora_up_rows  = lora_up->ne[ggml_n_dims(lora_up) - 1];
+    lora_up               = ggml_reshape_2d(ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
+    auto lora_down_n_dims = ggml_n_dims(lora_down);
+    // assume n_dims should always be a multiple of 2 (otherwise rank 1 doesn't work)
+    lora_down_n_dims       = (lora_down_n_dims + lora_down_n_dims % 2);
+    int64_t lora_down_rows = lora_down->ne[lora_down_n_dims - 1];
+    lora_down              = ggml_reshape_2d(ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
+
+    // ggml_mul_mat requires tensor b transposed
+    lora_down = ggml_cont(ctx, ggml_transpose(ctx, lora_down));
+    if (lora_mid == NULL) {
+        updown = ggml_mul_mat(ctx, lora_up, lora_down);
+        updown = ggml_cont(ctx, ggml_transpose(ctx, updown));
+    } else {
+        // undoing tucker decomposition for conv layers.
+        // lora_mid  has shape (3,    3,   Rank, Rank)
+        // lora_down has shape (Rank, In,  1,    1)
+        // lora_up   has shape (Rank, Out, 1,    1)
+        // conv layer shape is (3,    3,   Out,  In)
+        updown = ggml_mul_n_mode(ctx, ggml_mul_n_mode(ctx, lora_mid, lora_down, 3), lora_up, 2);
+        updown = ggml_cont(ctx, updown);
+    }
+    return updown;
+}
+
+// Kronecker product
+// [ne03,ne02,ne01,ne00] x [ne13,ne12,ne11,ne10] => [ne03*ne13,ne02*ne12,ne01*ne11,ne00*ne10]
+__STATIC_INLINE__ struct ggml_tensor* ggml_kronecker(ggml_context* ctx, struct ggml_tensor* a, struct ggml_tensor* b) {
+    return ggml_mul(ctx,
+                    ggml_upscale_ext(ctx,
+                                     a,
+                                     a->ne[0] * b->ne[0],
+                                     a->ne[1] * b->ne[1],
+                                     a->ne[2] * b->ne[2],
+                                     a->ne[3] * b->ne[3]),
+                    b);
+}
+
 __STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void* user_data) {
     (void)level;
     (void)user_data;
@@ -318,8 +383,10 @@ __STATIC_INLINE__ void sd_apply_mask(struct ggml_tensor* image_data,
     for (int ix = 0; ix < width; ix++) {
         for (int iy = 0; iy < height; iy++) {
             float m = ggml_tensor_get_f32(mask, ix, iy);
+            m       = round(m);  // inpaint models need binary masks
+            ggml_tensor_set_f32(mask, m, ix, iy);
             for (int k = 0; k < channels; k++) {
-                float value = ((float)(m < 254.5/255)) * (ggml_tensor_get_f32(image_data, ix, iy, k) - .5) + .5;
+                float value = (1 - m) * (ggml_tensor_get_f32(image_data, ix, iy, k) - .5) + .5;
                 ggml_tensor_set_f32(output, value, ix, iy, k);
             }
         }
@@ -987,8 +1054,8 @@ __STATIC_INLINE__ size_t ggml_tensor_num(ggml_context* ctx) {
 }
 
 /* SDXL with LoRA requires more space */
-#define MAX_PARAMS_TENSOR_NUM 15360
-#define MAX_GRAPH_SIZE 15360
+#define MAX_PARAMS_TENSOR_NUM 32768
+#define MAX_GRAPH_SIZE 32768
 
 struct GGMLRunner {
 protected: