feat: increase work_ctx memory buffer size (#814)

Thank you for your contribution.

.gitignore

@@ -1,10 +1,10 @@
 build*/
+cmake-build-*/
 test/
 .vscode/
+.idea/
 .cache/
 *.swp
-.vscode/
-.idea/
 *.bat
 *.bin
 *.exe

CMakeLists.txt

@@ -149,3 +149,7 @@ if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
+set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)

Dockerfile.sycl

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

README.md

@@ -334,9 +334,9 @@ arguments:
   --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])
   --skip-layer-start START           SLG enabling point: (default: 0.01)
   --skip-layer-end END               SLG disabling point: (default: 0.2)
-  --scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)
+  --scheduler {discrete, karras, exponential, ays, gits, smoothstep} Denoiser sigma scheduler (default: discrete)
   --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
-                                     sampling method (default: "euler_a")
+                                     sampling method (default: "euler" for Flux/SD3/Wan, "euler_a" otherwise)
   --steps  STEPS                     number of sample steps (default: 20)
   --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
@@ -347,7 +347,7 @@ arguments:
   --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])
   --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)
   --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)
-  --high-noise-scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)
+  --high-noise-scheduler {discrete, karras, exponential, ays, gits, smoothstep} Denoiser sigma scheduler (default: discrete)
   --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
                                      (high noise) sampling method (default: "euler_a")
   --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)

clip.hpp

@@ -548,9 +548,15 @@ protected:
     int64_t embed_dim;
     int64_t vocab_size;
     int64_t num_positions;
+    bool force_clip_f32;
 
     void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
-        enum ggml_type token_wtype    = GGML_TYPE_F32;
+        enum ggml_type token_wtype = GGML_TYPE_F32;
+        if (!force_clip_f32) {
+            auto tensor_type = tensor_types.find(prefix + "token_embedding.weight");
+            if (tensor_type != tensor_types.end())
+                token_wtype = tensor_type->second;
+        }
         enum ggml_type position_wtype = GGML_TYPE_F32;
 
         params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
@@ -560,10 +566,12 @@ protected:
 public:
     CLIPEmbeddings(int64_t embed_dim,
                    int64_t vocab_size    = 49408,
-                   int64_t num_positions = 77)
+                   int64_t num_positions = 77,
+                   bool force_clip_f32   = false)
         : embed_dim(embed_dim),
           vocab_size(vocab_size),
-          num_positions(num_positions) {
+          num_positions(num_positions),
+          force_clip_f32(force_clip_f32) {
     }
 
     struct ggml_tensor* get_token_embed_weight() {
@@ -678,12 +686,11 @@ public:
     int32_t n_head            = 12;
     int32_t n_layer           = 12;    // num_hidden_layers
     int32_t projection_dim    = 1280;  // only for OPEN_CLIP_VIT_BIGG_14
-    int32_t clip_skip         = -1;
     bool with_final_ln        = true;
 
     CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                   bool with_final_ln  = true,
-                  int clip_skip_value = -1)
+                  bool force_clip_f32 = false)
         : version(version), with_final_ln(with_final_ln) {
         if (version == OPEN_CLIP_VIT_H_14) {
             hidden_size       = 1024;
@@ -696,20 +703,12 @@ public:
             n_head            = 20;
             n_layer           = 32;
         }
-        set_clip_skip(clip_skip_value);
 
-        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token));
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token, force_clip_f32));
         blocks["encoder"]          = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
 
-    void set_clip_skip(int skip) {
-        if (skip <= 0) {
-            skip = -1;
-        }
-        clip_skip = skip;
-    }
-
     struct ggml_tensor* get_token_embed_weight() {
         auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         return embeddings->get_token_embed_weight();
@@ -720,7 +719,8 @@ public:
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* tkn_embeddings,
                                 size_t max_token_idx = 0,
-                                bool return_pooled   = false) {
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
         // input_ids: [N, n_token]
         auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
@@ -889,8 +889,8 @@ struct CLIPTextModelRunner : public GGMLRunner {
                         const std::string prefix,
                         CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                         bool with_final_ln  = true,
-                        int clip_skip_value = -1)
-        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, clip_skip_value) {
+                        bool force_clip_f32 = false)
+        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, force_clip_f32) {
         model.init(params_ctx, tensor_types, prefix);
     }
 
@@ -898,10 +898,6 @@ struct CLIPTextModelRunner : public GGMLRunner {
         return "clip";
     }
 
-    void set_clip_skip(int clip_skip) {
-        model.set_clip_skip(clip_skip);
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
@@ -911,7 +907,8 @@ struct CLIPTextModelRunner : public GGMLRunner {
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* embeddings,
                                 size_t max_token_idx = 0,
-                                bool return_pooled   = false) {
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
         if (input_ids->ne[0] > model.n_token) {
@@ -919,14 +916,15 @@ struct CLIPTextModelRunner : public GGMLRunner {
             input_ids = ggml_reshape_2d(ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
         }
 
-        return model.forward(ctx, backend, input_ids, embeddings, max_token_idx, return_pooled);
+        return model.forward(ctx, backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
     }
 
     struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
                                     int num_custom_embeddings    = 0,
                                     void* custom_embeddings_data = NULL,
                                     size_t max_token_idx         = 0,
-                                    bool return_pooled           = false) {
+                                    bool return_pooled           = false,
+                                    int clip_skip                = -1) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         input_ids = to_backend(input_ids);
@@ -945,7 +943,7 @@ struct CLIPTextModelRunner : public GGMLRunner {
             embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
         }
 
-        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled);
+        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -958,10 +956,11 @@ struct CLIPTextModelRunner : public GGMLRunner {
                  void* custom_embeddings_data,
                  size_t max_token_idx,
                  bool return_pooled,
+                 int clip_skip,
                  ggml_tensor** output,
                  ggml_context* output_ctx = NULL) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled);
+            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
         };
         GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
     }

conditioner.hpp

@@ -61,30 +61,16 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                       const String2GGMLType& tensor_types,
                                       const std::string& embd_dir,
                                       SDVersion version = VERSION_SD1,
-                                      PMVersion pv      = PM_VERSION_1,
-                                      int clip_skip     = -1)
+                                      PMVersion pv      = PM_VERSION_1)
         : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
+        bool force_clip_f32 = embd_dir.size() > 0;
         if (sd_version_is_sd1(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
         } else if (sd_version_is_sd2(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
         } else if (sd_version_is_sdxl(version)) {
-            text_model  = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
-        }
-        set_clip_skip(clip_skip);
-    }
-
-    void set_clip_skip(int clip_skip) {
-        if (clip_skip <= 0) {
-            clip_skip = 1;
-            if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) {
-                clip_skip = 2;
-            }
-        }
-        text_model->set_clip_skip(clip_skip);
-        if (sd_version_is_sdxl(version)) {
-            text_model2->set_clip_skip(clip_skip);
+            text_model  = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
+            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
         }
     }
 
@@ -129,7 +115,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             return true;
         }
         struct ggml_init_params params;
-        params.mem_size               = 10 * 1024 * 1024;  // max for custom embeddings 10 MB
+        params.mem_size               = 100 * 1024 * 1024;  // max for custom embeddings 100 MB
         params.mem_buffer             = NULL;
         params.no_alloc               = false;
         struct ggml_context* embd_ctx = ggml_init(params);
@@ -412,7 +398,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                              int height,
                                              int adm_in_channels  = -1,
                                              bool zero_out_masked = false) {
-        set_clip_skip(clip_skip);
         int64_t t0                               = ggml_time_ms();
         struct ggml_tensor* hidden_states        = NULL;  // [N, n_token, hidden_size]
         struct ggml_tensor* chunk_hidden_states  = NULL;  // [n_token, hidden_size] or [n_token, hidden_size + hidden_size2]
@@ -421,6 +406,10 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         struct ggml_tensor* pooled               = NULL;
         std::vector<float> hidden_states_vec;
 
+        if (clip_skip <= 0) {
+            clip_skip = (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) ? 2 : 1;
+        }
+
         size_t chunk_len   = 77;
         size_t chunk_count = tokens.size() / chunk_len;
         for (int chunk_idx = 0; chunk_idx < chunk_count; chunk_idx++) {
@@ -455,6 +444,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                     token_embed_custom.data(),
                                     max_token_idx,
                                     false,
+                                    clip_skip,
                                     &chunk_hidden_states1,
                                     work_ctx);
                 if (sd_version_is_sdxl(version)) {
@@ -464,6 +454,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                          token_embed_custom.data(),
                                          max_token_idx,
                                          false,
+                                         clip_skip,
                                          &chunk_hidden_states2, work_ctx);
                     // concat
                     chunk_hidden_states = ggml_tensor_concat(work_ctx, chunk_hidden_states1, chunk_hidden_states2, 0);
@@ -475,6 +466,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                              token_embed_custom.data(),
                                              max_token_idx,
                                              true,
+                                             clip_skip,
                                              &pooled,
                                              work_ctx);
                     }
@@ -669,21 +661,11 @@ struct SD3CLIPEmbedder : public Conditioner {
 
     SD3CLIPEmbedder(ggml_backend_t backend,
                     bool offload_params_to_cpu,
-                    const String2GGMLType& tensor_types = {},
-                    int clip_skip                       = -1)
+                    const String2GGMLType& tensor_types = {})
         : clip_g_tokenizer(0) {
         clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
         clip_g = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
         t5     = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.t5xxl.transformer");
-        set_clip_skip(clip_skip);
-    }
-
-    void set_clip_skip(int clip_skip) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l->set_clip_skip(clip_skip);
-        clip_g->set_clip_skip(clip_skip);
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
@@ -780,7 +762,6 @@ struct SD3CLIPEmbedder : public Conditioner {
                                              std::vector<std::pair<std::vector<int>, std::vector<float>>> token_and_weights,
                                              int clip_skip,
                                              bool zero_out_masked = false) {
-        set_clip_skip(clip_skip);
         auto& clip_l_tokens  = token_and_weights[0].first;
         auto& clip_l_weights = token_and_weights[0].second;
         auto& clip_g_tokens  = token_and_weights[1].first;
@@ -788,6 +769,10 @@ struct SD3CLIPEmbedder : public Conditioner {
         auto& t5_tokens      = token_and_weights[2].first;
         auto& t5_weights     = token_and_weights[2].second;
 
+        if (clip_skip <= 0) {
+            clip_skip = 2;
+        }
+
         int64_t t0                                 = ggml_time_ms();
         struct ggml_tensor* hidden_states          = NULL;  // [N, n_token*2, 4096]
         struct ggml_tensor* chunk_hidden_states    = NULL;  // [n_token*2, 4096]
@@ -818,6 +803,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                 NULL,
                                 max_token_idx,
                                 false,
+                                clip_skip,
                                 &chunk_hidden_states_l,
                                 work_ctx);
                 {
@@ -845,6 +831,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                     NULL,
                                     max_token_idx,
                                     true,
+                                    clip_skip,
                                     &pooled_l,
                                     work_ctx);
                 }
@@ -866,6 +853,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                 NULL,
                                 max_token_idx,
                                 false,
+                                clip_skip,
                                 &chunk_hidden_states_g,
                                 work_ctx);
 
@@ -894,6 +882,7 @@ struct SD3CLIPEmbedder : public Conditioner {
                                     NULL,
                                     max_token_idx,
                                     true,
+                                    clip_skip,
                                     &pooled_g,
                                     work_ctx);
                 }
@@ -1017,18 +1006,9 @@ struct FluxCLIPEmbedder : public Conditioner {
 
     FluxCLIPEmbedder(ggml_backend_t backend,
                      bool offload_params_to_cpu,
-                     const String2GGMLType& tensor_types = {},
-                     int clip_skip                       = -1) {
+                     const String2GGMLType& tensor_types = {}) {
         clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
         t5     = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.t5xxl.transformer");
-        set_clip_skip(clip_skip);
-    }
-
-    void set_clip_skip(int clip_skip) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l->set_clip_skip(clip_skip);
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
@@ -1109,12 +1089,15 @@ struct FluxCLIPEmbedder : public Conditioner {
                                              std::vector<std::pair<std::vector<int>, std::vector<float>>> token_and_weights,
                                              int clip_skip,
                                              bool zero_out_masked = false) {
-        set_clip_skip(clip_skip);
         auto& clip_l_tokens  = token_and_weights[0].first;
         auto& clip_l_weights = token_and_weights[0].second;
         auto& t5_tokens      = token_and_weights[1].first;
         auto& t5_weights     = token_and_weights[1].second;
 
+        if (clip_skip <= 0) {
+            clip_skip = 2;
+        }
+
         int64_t t0                              = ggml_time_ms();
         struct ggml_tensor* hidden_states       = NULL;  // [N, n_token, 4096]
         struct ggml_tensor* chunk_hidden_states = NULL;  // [n_token, 4096]
@@ -1143,6 +1126,7 @@ struct FluxCLIPEmbedder : public Conditioner {
                                 NULL,
                                 max_token_idx,
                                 true,
+                                clip_skip,
                                 &pooled,
                                 work_ctx);
             }
@@ -1241,7 +1225,6 @@ struct T5CLIPEmbedder : public Conditioner {
     T5CLIPEmbedder(ggml_backend_t backend,
                    bool offload_params_to_cpu,
                    const String2GGMLType& tensor_types = {},
-                   int clip_skip                       = -1,
                    bool use_mask                       = false,
                    int mask_pad                        = 1,
                    bool is_umt5                        = false)
@@ -1249,9 +1232,6 @@ struct T5CLIPEmbedder : public Conditioner {
         t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.t5xxl.transformer", is_umt5);
     }
 
-    void set_clip_skip(int clip_skip) {
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
         t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
     }

denoiser.hpp

@@ -251,6 +251,35 @@ struct KarrasSchedule : SigmaSchedule {
     }
 };
 
+// Close to Beta Schedule, but increadably simple in code.
+struct SmoothStepSchedule : SigmaSchedule {
+    static constexpr float smoothstep(float x) {
+        return x * x * (3.0f - 2.0f * x);
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, float /*sigma_min*/, float /*sigma_max*/, t_to_sigma_t t_to_sigma) override {
+        std::vector<float> result;
+        result.reserve(n + 1);
+
+        const int t_max = TIMESTEPS - 1;
+        if (n == 0) {
+            return result;
+        } else if (n == 1) {
+            result.push_back(t_to_sigma((float)t_max));
+            result.push_back(0.f);
+            return result;
+        }
+
+        for (uint32_t i = 0; i < n; i++) {
+            float u = 1.f - float(i) / float(n);
+            result.push_back(t_to_sigma(std::round(smoothstep(u) * t_max)));
+        }
+
+        result.push_back(0.f);
+        return result;
+    }
+};
+
 struct Denoiser {
     std::shared_ptr<SigmaSchedule> scheduler                                                 = std::make_shared<DiscreteSchedule>();
     virtual float sigma_min()                                                                = 0;

diffusion_model.hpp

@@ -95,8 +95,9 @@ struct MMDiTModel : public DiffusionModel {
 
     MMDiTModel(ggml_backend_t backend,
                bool offload_params_to_cpu,
+               bool flash_attn                     = false,
                const String2GGMLType& tensor_types = {})
-        : mmdit(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model") {
+        : mmdit(backend, offload_params_to_cpu, flash_attn, tensor_types, "model.diffusion_model") {
     }
 
     std::string get_desc() {

examples/cli/main.cpp

@@ -238,9 +238,9 @@ void print_usage(int argc, const char* argv[]) {
     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");
-    printf("  --scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)\n");
+    printf("  --scheduler {discrete, karras, exponential, ays, gits, smoothstep} Denoiser sigma scheduler (default: discrete)\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("                                     sampling method (default: \"euler\" for Flux/SD3/Wan, \"euler_a\" otherwise)\n");
     printf("  --steps  STEPS                     number of sample steps (default: 20)\n");
     printf("  --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)\n");
     printf("  --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)\n");
@@ -251,7 +251,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])\n");
     printf("  --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)\n");
     printf("  --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)\n");
-    printf("  --high-noise-scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)\n");
+    printf("  --high-noise-scheduler {discrete, karras, exponential, ays, gits, smoothstep} Denoiser sigma scheduler (default: discrete)\n");
     printf("  --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}\n");
     printf("                                     (high noise) sampling method (default: \"euler_a\")\n");
     printf("  --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)\n");
@@ -1202,6 +1202,10 @@ int main(int argc, const char* argv[]) {
         return 1;
     }
 
+    if (params.sample_params.sample_method == SAMPLE_METHOD_DEFAULT) {
+        params.sample_params.sample_method = sd_get_default_sample_method(sd_ctx);
+    }
+
     sd_image_t* results;
     int num_results = 1;
     if (params.mode == IMG_GEN) {

ggml_extend.hpp

@@ -56,6 +56,25 @@
 #define __STATIC_INLINE__ static inline
 #endif
 
+__STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void*) {
+    switch (level) {
+        case GGML_LOG_LEVEL_DEBUG:
+            LOG_DEBUG(text);
+            break;
+        case GGML_LOG_LEVEL_INFO:
+            LOG_INFO(text);
+            break;
+        case GGML_LOG_LEVEL_WARN:
+            LOG_WARN(text);
+            break;
+        case GGML_LOG_LEVEL_ERROR:
+            LOG_ERROR(text);
+            break;
+        default:
+            LOG_DEBUG(text);
+    }
+}
+
 static_assert(GGML_MAX_NAME >= 128, "GGML_MAX_NAME must be at least 128");
 
 // n-mode trensor-matrix product
@@ -124,13 +143,6 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_kronecker(ggml_context* ctx, struct g
                     b);
 }
 
-__STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void* user_data) {
-    (void)level;
-    (void)user_data;
-    fputs(text, stderr);
-    fflush(stderr);
-}
-
 __STATIC_INLINE__ void ggml_tensor_set_f32_randn(struct ggml_tensor* tensor, std::shared_ptr<RNG> rng) {
     uint32_t n                        = (uint32_t)ggml_nelements(tensor);
     std::vector<float> random_numbers = rng->randn(n);

mmdit.hpp

@@ -147,14 +147,16 @@ public:
     int64_t num_heads;
     bool pre_only;
     std::string qk_norm;
+    bool flash_attn;
 
 public:
     SelfAttention(int64_t dim,
                   int64_t num_heads   = 8,
                   std::string qk_norm = "",
                   bool qkv_bias       = false,
-                  bool pre_only       = false)
-        : num_heads(num_heads), pre_only(pre_only), qk_norm(qk_norm) {
+                  bool pre_only       = false,
+                  bool flash_attn     = false)
+        : num_heads(num_heads), pre_only(pre_only), qk_norm(qk_norm), flash_attn(flash_attn) {
         int64_t d_head = dim / num_heads;
         blocks["qkv"]  = std::shared_ptr<GGMLBlock>(new Linear(dim, dim * 3, qkv_bias));
         if (!pre_only) {
@@ -206,8 +208,8 @@ public:
                                 ggml_backend_t backend,
                                 struct ggml_tensor* x) {
         auto qkv = pre_attention(ctx, x);
-        x        = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
-        x        = post_attention(ctx, x);                                                  // [N, n_token, dim]
+        x        = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, NULL, false, false, true);  // [N, n_token, dim]
+        x        = post_attention(ctx, x);                                                                            // [N, n_token, dim]
         return x;
     }
 };
@@ -232,6 +234,7 @@ public:
     int64_t num_heads;
     bool pre_only;
     bool self_attn;
+    bool flash_attn;
 
 public:
     DismantledBlock(int64_t hidden_size,
@@ -240,16 +243,17 @@ public:
                     std::string qk_norm = "",
                     bool qkv_bias       = false,
                     bool pre_only       = false,
-                    bool self_attn      = false)
+                    bool self_attn      = false,
+                    bool flash_attn     = false)
         : num_heads(num_heads), pre_only(pre_only), self_attn(self_attn) {
         // rmsnorm is always Flase
         // scale_mod_only is always Flase
         // swiglu is always Flase
         blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
-        blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only));
+        blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only, flash_attn));
 
         if (self_attn) {
-            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false));
+            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false, flash_attn));
         }
 
         if (!pre_only) {
@@ -435,8 +439,8 @@ public:
             auto qkv2          = std::get<1>(qkv_intermediates);
             auto intermediates = std::get<2>(qkv_intermediates);
 
-            auto attn_out  = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads);     // [N, n_token, dim]
-            auto attn2_out = ggml_nn_attention_ext(ctx, backend, qkv2[0], qkv2[1], qkv2[2], num_heads);  // [N, n_token, dim]
+            auto attn_out  = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, NULL, false, false, flash_attn);     // [N, n_token, dim]
+            auto attn2_out = ggml_nn_attention_ext(ctx, backend, qkv2[0], qkv2[1], qkv2[2], num_heads, NULL, false, false, flash_attn);  // [N, n_token, dim]
             x              = post_attention_x(ctx,
                                               attn_out,
                                               attn2_out,
@@ -452,7 +456,7 @@ public:
             auto qkv               = qkv_intermediates.first;
             auto intermediates     = qkv_intermediates.second;
 
-            auto attn_out = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
+            auto attn_out = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, NULL, false, false, flash_attn);  // [N, n_token, dim]
             x             = post_attention(ctx,
                                            attn_out,
                                            intermediates[0],
@@ -468,6 +472,7 @@ public:
 __STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*>
 block_mixing(struct ggml_context* ctx,
              ggml_backend_t backend,
+             bool flash_attn,
              struct ggml_tensor* context,
              struct ggml_tensor* x,
              struct ggml_tensor* c,
@@ -497,8 +502,8 @@ block_mixing(struct ggml_context* ctx,
         qkv.push_back(ggml_concat(ctx, context_qkv[i], x_qkv[i], 1));
     }
 
-    auto attn         = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], x_block->num_heads);  // [N, n_context + n_token, hidden_size]
-    attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                              // [n_context + n_token, N, hidden_size]
+    auto attn         = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], x_block->num_heads, NULL, false, false, flash_attn);  // [N, n_context + n_token, hidden_size]
+    attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                                                              // [n_context + n_token, N, hidden_size]
     auto context_attn = ggml_view_3d(ctx,
                                      attn,
                                      attn->ne[0],
@@ -556,6 +561,8 @@ block_mixing(struct ggml_context* ctx,
 }
 
 struct JointBlock : public GGMLBlock {
+    bool flash_attn;
+
 public:
     JointBlock(int64_t hidden_size,
                int64_t num_heads,
@@ -563,9 +570,11 @@ public:
                std::string qk_norm = "",
                bool qkv_bias       = false,
                bool pre_only       = false,
-               bool self_attn_x    = false) {
-        blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only));
-        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x));
+               bool self_attn_x    = false,
+               bool flash_attn     = false)
+        : flash_attn(flash_attn) {
+        blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only, false, flash_attn));
+        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x, flash_attn));
     }
 
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
@@ -576,7 +585,7 @@ public:
         auto context_block = std::dynamic_pointer_cast<DismantledBlock>(blocks["context_block"]);
         auto x_block       = std::dynamic_pointer_cast<DismantledBlock>(blocks["x_block"]);
 
-        return block_mixing(ctx, backend, context, x, c, context_block, x_block);
+        return block_mixing(ctx, backend, flash_attn, context, x, c, context_block, x_block);
     }
 };
 
@@ -634,6 +643,7 @@ protected:
     int64_t context_embedder_out_dim = 1536;
     int64_t hidden_size;
     std::string qk_norm;
+    bool flash_attn = false;
 
     void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
         enum ggml_type wtype = GGML_TYPE_F32;
@@ -641,7 +651,8 @@ protected:
     }
 
 public:
-    MMDiT(const String2GGMLType& tensor_types = {}) {
+    MMDiT(bool flash_attn = false, const String2GGMLType& tensor_types = {})
+        : flash_attn(flash_attn) {
         // input_size is always None
         // learn_sigma is always False
         // register_length is alwalys 0
@@ -709,7 +720,8 @@ public:
                                                                                                     qk_norm,
                                                                                                     true,
                                                                                                     i == depth - 1,
-                                                                                                    i <= d_self));
+                                                                                                    i <= d_self,
+                                                                                                    flash_attn));
         }
 
         blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
@@ -856,9 +868,10 @@ struct MMDiTRunner : public GGMLRunner {
 
     MMDiTRunner(ggml_backend_t backend,
                 bool offload_params_to_cpu,
+                bool flash_attn,
                 const String2GGMLType& tensor_types = {},
                 const std::string prefix            = "")
-        : GGMLRunner(backend, offload_params_to_cpu), mmdit(tensor_types) {
+        : GGMLRunner(backend, offload_params_to_cpu), mmdit(flash_attn, tensor_types) {
         mmdit.init(params_ctx, tensor_types, prefix);
     }
 
@@ -957,7 +970,7 @@ struct MMDiTRunner : public GGMLRunner {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
         ggml_backend_t backend             = ggml_backend_cpu_init();
         ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend, false));
+        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend, false, false));
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 

model.cpp

@@ -107,7 +107,7 @@ const char* unused_tensors[] = {
 };
 
 bool is_unused_tensor(std::string name) {
-    for (int i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
+    for (size_t i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
         if (starts_with(name, unused_tensors[i])) {
             return true;
         }
@@ -2310,7 +2310,7 @@ std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std
         if (type_name == "f32") {
             tensor_type = GGML_TYPE_F32;
         } else {
-            for (size_t i = 0; i < SD_TYPE_COUNT; i++) {
+            for (size_t i = 0; i < GGML_TYPE_COUNT; i++) {
                 auto trait = ggml_get_type_traits((ggml_type)i);
                 if (trait->to_float && trait->type_size && type_name == trait->type_name) {
                     tensor_type = (ggml_type)i;

model.h

@@ -119,7 +119,7 @@ struct TensorStorage {
 
     size_t file_index = 0;
     int index_in_zip  = -1;  // >= means stored in a zip file
-    size_t offset     = 0;   // offset in file
+    uint64_t offset   = 0;   // offset in file
 
     TensorStorage() = default;
 
@@ -164,10 +164,10 @@ struct TensorStorage {
 
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
-        size_t chunk_size = nbytes_to_read() / n;
+        uint64_t chunk_size = nbytes_to_read() / n;
         // printf("%d/%d\n", chunk_size, nbytes_to_read());
         reverse_ne();
-        for (int i = 0; i < n; i++) {
+        for (size_t i = 0; i < n; i++) {
             TensorStorage chunk_i = *this;
             chunk_i.ne[0]         = ne[0] / n;
             chunk_i.offset        = offset + i * chunk_size;

preprocessing.hpp

@@ -164,7 +164,7 @@ void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float lo
 
 uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
     struct ggml_init_params params;
-    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10
+    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10MB
     params.mem_buffer             = NULL;
     params.no_alloc               = false;
     struct ggml_context* work_ctx = ggml_init(params);

stable-diffusion.cpp

@@ -43,7 +43,7 @@ const char* model_version_to_str[] = {
 };
 
 const char* sampling_methods_str[] = {
-    "Euler A",
+    "default",
     "Euler",
     "Heun",
     "DPM2",
@@ -55,6 +55,7 @@ const char* sampling_methods_str[] = {
     "LCM",
     "DDIM \"trailing\"",
     "TCD",
+    "Euler A",
 };
 
 /*================================================== Helper Functions ================================================*/
@@ -145,7 +146,6 @@ public:
 #endif
 #ifdef SD_USE_METAL
         LOG_DEBUG("Using Metal backend");
-        ggml_log_set(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
 #ifdef SD_USE_VULKAN
@@ -192,6 +192,8 @@ public:
             rng = std::make_shared<PhiloxRNG>();
         }
 
+        ggml_log_set(ggml_log_callback_default, nullptr);
+
         init_backend();
 
         ModelLoader model_loader;
@@ -264,7 +266,9 @@ public:
         }
 
         LOG_INFO("Version: %s ", model_version_to_str[version]);
-        ggml_type wtype = (ggml_type)sd_ctx_params->wtype;
+        ggml_type wtype = (int)sd_ctx_params->wtype < std::min<int>(SD_TYPE_COUNT, GGML_TYPE_COUNT)
+                              ? (ggml_type)sd_ctx_params->wtype
+                              : GGML_TYPE_COUNT;
         if (wtype == GGML_TYPE_COUNT) {
             model_wtype = model_loader.get_sd_wtype();
             if (model_wtype == GGML_TYPE_COUNT) {
@@ -292,11 +296,6 @@ public:
             model_loader.set_wtype_override(wtype);
         }
 
-        if (sd_version_is_sdxl(version)) {
-            vae_wtype = GGML_TYPE_F32;
-            model_loader.set_wtype_override(GGML_TYPE_F32, "vae.");
-        }
-
         LOG_INFO("Weight type:                 %s", ggml_type_name(model_wtype));
         LOG_INFO("Conditioner weight type:     %s", ggml_type_name(conditioner_wtype));
         LOG_INFO("Diffusion model weight type: %s", ggml_type_name(diffusion_model_wtype));
@@ -344,14 +343,12 @@ public:
                 LOG_INFO("Using flash attention in the diffusion model");
             }
             if (sd_version_is_sd3(version)) {
-                if (sd_ctx_params->diffusion_flash_attn) {
-                    LOG_WARN("flash attention in this diffusion model is currently unsupported!");
-                }
                 cond_stage_model = std::make_shared<SD3CLIPEmbedder>(clip_backend,
                                                                      offload_params_to_cpu,
                                                                      model_loader.tensor_storages_types);
                 diffusion_model  = std::make_shared<MMDiTModel>(backend,
                                                                offload_params_to_cpu,
+                                                               sd_ctx_params->diffusion_flash_attn,
                                                                model_loader.tensor_storages_types);
             } else if (sd_version_is_flux(version)) {
                 bool is_chroma = false;
@@ -362,10 +359,18 @@ public:
                     }
                 }
                 if (is_chroma) {
+                    if (sd_ctx_params->diffusion_flash_attn && sd_ctx_params->chroma_use_dit_mask) {
+                        LOG_WARN(
+                            "!!!It looks like you are using Chroma with flash attention. "
+                            "This is currently unsupported. "
+                            "If you find that the generated images are broken, "
+                            "try either disabling flash attention or specifying "
+                            "--chroma-disable-dit-mask as a workaround.");
+                    }
+
                     cond_stage_model = std::make_shared<T5CLIPEmbedder>(clip_backend,
                                                                         offload_params_to_cpu,
                                                                         model_loader.tensor_storages_types,
-                                                                        -1,
                                                                         sd_ctx_params->chroma_use_t5_mask,
                                                                         sd_ctx_params->chroma_t5_mask_pad);
                 } else {
@@ -383,7 +388,6 @@ public:
                 cond_stage_model = std::make_shared<T5CLIPEmbedder>(clip_backend,
                                                                     offload_params_to_cpu,
                                                                     model_loader.tensor_storages_types,
-                                                                    -1,
                                                                     true,
                                                                     1,
                                                                     true);
@@ -744,6 +748,10 @@ public:
                 denoiser->scheduler          = std::make_shared<GITSSchedule>();
                 denoiser->scheduler->version = version;
                 break;
+            case SMOOTHSTEP:
+                LOG_INFO("Running with SmoothStep scheduler");
+                denoiser->scheduler = std::make_shared<SmoothStepSchedule>();
+                break;
             case DEFAULT:
                 // Don't touch anything.
                 break;
@@ -1455,11 +1463,14 @@ public:
 #define NONE_STR "NONE"
 
 const char* sd_type_name(enum sd_type_t type) {
-    return ggml_type_name((ggml_type)type);
+    if ((int)type < std::min<int>(SD_TYPE_COUNT, GGML_TYPE_COUNT)) {
+        return ggml_type_name((ggml_type)type);
+    }
+    return NONE_STR;
 }
 
 enum sd_type_t str_to_sd_type(const char* str) {
-    for (int i = 0; i < SD_TYPE_COUNT; i++) {
+    for (int i = 0; i < std::min<int>(SD_TYPE_COUNT, GGML_TYPE_COUNT); i++) {
         auto trait = ggml_get_type_traits((ggml_type)i);
         if (!strcmp(str, trait->type_name)) {
             return (enum sd_type_t)i;
@@ -1490,7 +1501,7 @@ enum rng_type_t str_to_rng_type(const char* str) {
 }
 
 const char* sample_method_to_str[] = {
-    "euler_a",
+    "default",
     "euler",
     "heun",
     "dpm2",
@@ -1502,6 +1513,7 @@ const char* sample_method_to_str[] = {
     "lcm",
     "ddim_trailing",
     "tcd",
+    "euler_a",
 };
 
 const char* sd_sample_method_name(enum sample_method_t sample_method) {
@@ -1527,6 +1539,7 @@ const char* schedule_to_str[] = {
     "exponential",
     "ays",
     "gits",
+    "smoothstep",
 };
 
 const char* sd_schedule_name(enum scheduler_t scheduler) {
@@ -1546,7 +1559,7 @@ enum scheduler_t str_to_schedule(const char* str) {
 }
 
 void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
-    memset((void*)sd_ctx_params, 0, sizeof(sd_ctx_params_t));
+    *sd_ctx_params                         = {};
     sd_ctx_params->vae_decode_only         = true;
     sd_ctx_params->vae_tiling              = false;
     sd_ctx_params->free_params_immediately = true;
@@ -1630,6 +1643,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
 }
 
 void sd_sample_params_init(sd_sample_params_t* sample_params) {
+    *sample_params                             = {};
     sample_params->guidance.txt_cfg            = 7.0f;
     sample_params->guidance.img_cfg            = INFINITY;
     sample_params->guidance.distilled_guidance = 3.5f;
@@ -1638,7 +1652,7 @@ void sd_sample_params_init(sd_sample_params_t* sample_params) {
     sample_params->guidance.slg.layer_end      = 0.2f;
     sample_params->guidance.slg.scale          = 0.f;
     sample_params->scheduler                   = DEFAULT;
-    sample_params->sample_method               = EULER_A;
+    sample_params->sample_method               = SAMPLE_METHOD_DEFAULT;
     sample_params->sample_steps                = 20;
 }
 
@@ -1676,9 +1690,9 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
 }
 
 void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params) {
-    memset((void*)sd_img_gen_params, 0, sizeof(sd_img_gen_params_t));
-    sd_img_gen_params->clip_skip = -1;
+    *sd_img_gen_params = {};
     sd_sample_params_init(&sd_img_gen_params->sample_params);
+    sd_img_gen_params->clip_skip        = -1;
     sd_img_gen_params->ref_images_count = 0;
     sd_img_gen_params->width            = 512;
     sd_img_gen_params->height           = 512;
@@ -1735,7 +1749,7 @@ char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
 }
 
 void sd_vid_gen_params_init(sd_vid_gen_params_t* sd_vid_gen_params) {
-    memset((void*)sd_vid_gen_params, 0, sizeof(sd_vid_gen_params_t));
+    *sd_vid_gen_params = {};
     sd_sample_params_init(&sd_vid_gen_params->sample_params);
     sd_sample_params_init(&sd_vid_gen_params->high_noise_sample_params);
     sd_vid_gen_params->high_noise_sample_params.sample_steps = -1;
@@ -1759,6 +1773,7 @@ sd_ctx_t* new_sd_ctx(const sd_ctx_params_t* sd_ctx_params) {
 
     sd_ctx->sd = new StableDiffusionGGML();
     if (sd_ctx->sd == NULL) {
+        free(sd_ctx);
         return NULL;
     }
 
@@ -1779,6 +1794,17 @@ void free_sd_ctx(sd_ctx_t* sd_ctx) {
     free(sd_ctx);
 }
 
+enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx) {
+    if (sd_ctx != NULL && sd_ctx->sd != NULL) {
+        SDVersion version = sd_ctx->sd->version;
+        if (sd_version_is_dit(version))
+            return EULER;
+        else
+            return EULER_A;
+    }
+    return SAMPLE_METHOD_COUNT;
+}
+
 sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
                                     struct ggml_context* work_ctx,
                                     ggml_tensor* init_latent,
@@ -2170,19 +2196,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
     }
 
     struct ggml_init_params params;
-    params.mem_size = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-    if (sd_version_is_sd3(sd_ctx->sd->version)) {
-        params.mem_size *= 3;
-    }
-    if (sd_version_is_flux(sd_ctx->sd->version)) {
-        params.mem_size *= 4;
-    }
-    if (sd_ctx->sd->stacked_id) {
-        params.mem_size += static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-    }
-    params.mem_size += width * height * 3 * sizeof(float) * 3;
-    params.mem_size += width * height * 3 * sizeof(float) * 3 * sd_img_gen_params->ref_images_count;
-    params.mem_size *= sd_img_gen_params->batch_count;
+    params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
     params.mem_buffer = NULL;
     params.no_alloc   = false;
     // LOG_DEBUG("mem_size %u ", params.mem_size);
@@ -2343,6 +2357,11 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
         LOG_INFO("encode_first_stage completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
     }
 
+    enum sample_method_t sample_method = sd_img_gen_params->sample_params.sample_method;
+    if (sample_method == SAMPLE_METHOD_DEFAULT) {
+        sample_method = sd_get_default_sample_method(sd_ctx);
+    }
+
     sd_image_t* result_images = generate_image_internal(sd_ctx,
                                                         work_ctx,
                                                         init_latent,
@@ -2353,7 +2372,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                                                         sd_img_gen_params->sample_params.eta,
                                                         width,
                                                         height,
-                                                        sd_img_gen_params->sample_params.sample_method,
+                                                        sample_method,
                                                         sigmas,
                                                         seed,
                                                         sd_img_gen_params->batch_count,
@@ -2417,8 +2436,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
     }
 
     struct ggml_init_params params;
-    params.mem_size = static_cast<size_t>(200 * 1024) * 1024;  // 200 MB
-    params.mem_size += width * height * frames * 3 * sizeof(float) * 2;
+    params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
     params.mem_buffer = NULL;
     params.no_alloc   = false;
     // LOG_DEBUG("mem_size %u ", params.mem_size);

stable-diffusion.h

@@ -35,7 +35,7 @@ enum rng_type_t {
 };
 
 enum sample_method_t {
-    EULER_A,
+    SAMPLE_METHOD_DEFAULT,
     EULER,
     HEUN,
     DPM2,
@@ -47,6 +47,7 @@ enum sample_method_t {
     LCM,
     DDIM_TRAILING,
     TCD,
+    EULER_A,
     SAMPLE_METHOD_COUNT
 };
 
@@ -57,6 +58,7 @@ enum scheduler_t {
     EXPONENTIAL,
     AYS,
     GITS,
+    SMOOTHSTEP,
     SCHEDULE_COUNT
 };
 
@@ -237,6 +239,7 @@ SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);
 
 SD_API sd_ctx_t* new_sd_ctx(const sd_ctx_params_t* sd_ctx_params);
 SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
+SD_API enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx);
 
 SD_API void sd_sample_params_init(sd_sample_params_t* sample_params);
 SD_API char* sd_sample_params_to_str(const sd_sample_params_t* sample_params);

upscaler.cpp

@@ -19,13 +19,13 @@ struct UpscalerGGML {
 
     bool load_from_file(const std::string& esrgan_path,
                         bool offload_params_to_cpu) {
+        ggml_log_set(ggml_log_callback_default, nullptr);
 #ifdef SD_USE_CUDA
         LOG_DEBUG("Using CUDA backend");
         backend = ggml_backend_cuda_init(0);
 #endif
 #ifdef SD_USE_METAL
         LOG_DEBUG("Using Metal backend");
-        ggml_log_set(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
 #ifdef SD_USE_VULKAN
@@ -69,8 +69,7 @@ struct UpscalerGGML {
                  input_image.width, input_image.height, output_width, output_height);
 
         struct ggml_init_params params;
-        params.mem_size = output_width * output_height * 3 * sizeof(float) * 2;
-        params.mem_size += 2 * ggml_tensor_overhead();
+        params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
         params.mem_buffer = NULL;
         params.no_alloc   = false;
 
@@ -80,7 +79,7 @@ struct UpscalerGGML {
             LOG_ERROR("ggml_init() failed");
             return upscaled_image;
         }
-        LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
+        // LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
         ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
         sd_image_to_tensor(input_image.data, input_image_tensor);
 

util.cpp

@@ -414,7 +414,10 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
     if (written >= 0 && written < LOG_BUFFER_SIZE) {
         vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
     }
-    strncat(log_buffer, "\n", LOG_BUFFER_SIZE - strlen(log_buffer));
+    size_t len = strlen(log_buffer);
+    if (log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
 
     if (sd_log_cb) {
         sd_log_cb(level, log_buffer, sd_log_cb_data);

vae.hpp

@@ -588,7 +588,7 @@ struct AutoEncoderKL : public VAE {
         };
         // ggml_set_f32(z, 0.5f);
         // print_ggml_tensor(z);
-        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
     }
 
     void test() {

wan.hpp

@@ -1219,7 +1219,7 @@ namespace WAN {
 
         void test() {
             struct ggml_init_params params;
-            params.mem_size   = static_cast<size_t>(1000 * 1024 * 1024);  // 10 MB
+            params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
             params.mem_buffer = NULL;
             params.no_alloc   = false;