fix: avoid the issue of NaN for qwen-image on certain devices (#1249)

conditioner.hpp

@@ -34,6 +34,7 @@ struct Conditioner {
     virtual void free_params_buffer()                                                      = 0;
     virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors)    = 0;
     virtual size_t get_params_buffer_size()                                                = 0;
+    virtual void set_flash_attention_enabled(bool enabled)                                 = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
                                                                                           int n_threads,
@@ -115,6 +116,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return buffer_size;
     }
 
+    void set_flash_attention_enabled(bool enabled) override {
+        text_model->set_flash_attention_enabled(enabled);
+        if (sd_version_is_sdxl(version)) {
+            text_model2->set_flash_attention_enabled(enabled);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         text_model->set_weight_adapter(adapter);
         if (sd_version_is_sdxl(version)) {
@@ -783,6 +791,18 @@ struct SD3CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_flash_attention_enabled(bool enabled) override {
+        if (clip_l) {
+            clip_l->set_flash_attention_enabled(enabled);
+        }
+        if (clip_g) {
+            clip_g->set_flash_attention_enabled(enabled);
+        }
+        if (t5) {
+            t5->set_flash_attention_enabled(enabled);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (clip_l) {
             clip_l->set_weight_adapter(adapter);
@@ -1191,6 +1211,15 @@ struct FluxCLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_flash_attention_enabled(bool enabled) override {
+        if (clip_l) {
+            clip_l->set_flash_attention_enabled(enabled);
+        }
+        if (t5) {
+            t5->set_flash_attention_enabled(enabled);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {
         if (clip_l) {
             clip_l->set_weight_adapter(adapter);
@@ -1440,6 +1469,12 @@ struct T5CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_flash_attention_enabled(bool enabled) override {
+        if (t5) {
+            t5->set_flash_attention_enabled(enabled);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (t5) {
             t5->set_weight_adapter(adapter);
@@ -1650,6 +1685,10 @@ struct LLMEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_flash_attention_enabled(bool enabled) override {
+        llm->set_flash_attention_enabled(enabled);
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (llm) {
             llm->set_weight_adapter(adapter);

denoiser.hpp

@@ -1,6 +1,8 @@
 #ifndef __DENOISER_HPP__
 #define __DENOISER_HPP__
 
+#include <cmath>
+
 #include "ggml_extend.hpp"
 #include "gits_noise.inl"
 
@@ -351,6 +353,95 @@ struct SmoothStepScheduler : SigmaScheduler {
     }
 };
 
+struct BongTangentScheduler : SigmaScheduler {
+    static constexpr float kPi = 3.14159265358979323846f;
+
+    static std::vector<float> get_bong_tangent_sigmas(int steps, float slope, float pivot, float start, float end) {
+        std::vector<float> sigmas;
+        if (steps <= 0) {
+            return sigmas;
+        }
+
+        float smax   = ((2.0f / kPi) * atanf(-slope * (0.0f - pivot)) + 1.0f) * 0.5f;
+        float smin   = ((2.0f / kPi) * atanf(-slope * ((float)(steps - 1) - pivot)) + 1.0f) * 0.5f;
+        float srange = smax - smin;
+        float sscale = start - end;
+
+        sigmas.reserve(steps);
+
+        if (fabsf(srange) < 1e-8f) {
+            if (steps == 1) {
+                sigmas.push_back(start);
+                return sigmas;
+            }
+            for (int i = 0; i < steps; ++i) {
+                float t = (float)i / (float)(steps - 1);
+                sigmas.push_back(start + (end - start) * t);
+            }
+            return sigmas;
+        }
+
+        float inv_srange = 1.0f / srange;
+        for (int x = 0; x < steps; ++x) {
+            float v     = ((2.0f / kPi) * atanf(-slope * ((float)x - pivot)) + 1.0f) * 0.5f;
+            float sigma = ((v - smin) * inv_srange) * sscale + end;
+            sigmas.push_back(sigma);
+        }
+
+        return sigmas;
+    }
+
+    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;
+        if (n == 0) {
+            return result;
+        }
+
+        float start  = sigma_max;
+        float end    = sigma_min;
+        float middle = sigma_min + (sigma_max - sigma_min) * 0.5f;
+
+        float pivot_1 = 0.6f;
+        float pivot_2 = 0.6f;
+        float slope_1 = 0.2f;
+        float slope_2 = 0.2f;
+
+        int steps     = static_cast<int>(n) + 2;
+        int midpoint  = static_cast<int>(((float)steps * pivot_1 + (float)steps * pivot_2) * 0.5f);
+        int pivot_1_i = static_cast<int>((float)steps * pivot_1);
+        int pivot_2_i = static_cast<int>((float)steps * pivot_2);
+
+        float slope_scale = (float)steps / 40.0f;
+        slope_1           = slope_1 / slope_scale;
+        slope_2           = slope_2 / slope_scale;
+
+        int stage_2_len = steps - midpoint;
+        int stage_1_len = steps - stage_2_len;
+
+        std::vector<float> sigmas_1 = get_bong_tangent_sigmas(stage_1_len, slope_1, (float)pivot_1_i, start, middle);
+        std::vector<float> sigmas_2 = get_bong_tangent_sigmas(stage_2_len, slope_2, (float)(pivot_2_i - stage_1_len), middle, end);
+
+        if (!sigmas_1.empty()) {
+            sigmas_1.pop_back();
+        }
+
+        result.reserve(n + 1);
+        result.insert(result.end(), sigmas_1.begin(), sigmas_1.end());
+        result.insert(result.end(), sigmas_2.begin(), sigmas_2.end());
+
+        if (result.size() < n + 1) {
+            while (result.size() < n + 1) {
+                result.push_back(end);
+            }
+        } else if (result.size() > n + 1) {
+            result.resize(n + 1);
+        }
+
+        result[n] = 0.0f;
+        return result;
+    }
+};
+
 struct KLOptimalScheduler : SigmaScheduler {
     std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         std::vector<float> sigmas;
@@ -431,6 +522,10 @@ struct Denoiser {
                 LOG_INFO("get_sigmas with SmoothStep scheduler");
                 scheduler = std::make_shared<SmoothStepScheduler>();
                 break;
+            case BONG_TANGENT_SCHEDULER:
+                LOG_INFO("get_sigmas with bong_tangent scheduler");
+                scheduler = std::make_shared<BongTangentScheduler>();
+                break;
             case KL_OPTIMAL_SCHEDULER:
                 LOG_INFO("get_sigmas with KL Optimal scheduler");
                 scheduler = std::make_shared<KLOptimalScheduler>();
@@ -1634,6 +1729,216 @@ static bool sample_k_diffusion(sample_method_t method,
                 }
             }
         } break;
+        case RES_MULTISTEP_SAMPLE_METHOD:  // Res Multistep sampler
+        {
+            struct ggml_tensor* noise        = ggml_dup_tensor(work_ctx, x);
+            struct ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
+
+            bool have_old_sigma  = false;
+            float old_sigma_down = 0.0f;
+
+            auto t_fn     = [](float sigma) -> float { return -logf(sigma); };
+            auto sigma_fn = [](float t) -> float { return expf(-t); };
+            auto phi1_fn  = [](float t) -> float {
+                if (fabsf(t) < 1e-6f) {
+                    return 1.0f + t * 0.5f + (t * t) / 6.0f;
+                }
+                return (expf(t) - 1.0f) / t;
+            };
+            auto phi2_fn = [&](float t) -> float {
+                if (fabsf(t) < 1e-6f) {
+                    return 0.5f + t / 6.0f + (t * t) / 24.0f;
+                }
+                float phi1_val = phi1_fn(t);
+                return (phi1_val - 1.0f) / t;
+            };
+
+            for (int i = 0; i < steps; i++) {
+                ggml_tensor* denoised = model(x, sigmas[i], i + 1);
+                if (denoised == nullptr) {
+                    return false;
+                }
+
+                float sigma_from = sigmas[i];
+                float sigma_to   = sigmas[i + 1];
+                float sigma_up   = 0.0f;
+                float sigma_down = sigma_to;
+
+                if (eta > 0.0f) {
+                    float sigma_from_sq = sigma_from * sigma_from;
+                    float sigma_to_sq   = sigma_to * sigma_to;
+                    if (sigma_from_sq > 0.0f) {
+                        float term = sigma_to_sq * (sigma_from_sq - sigma_to_sq) / sigma_from_sq;
+                        if (term > 0.0f) {
+                            sigma_up = eta * std::sqrt(term);
+                        }
+                    }
+                    sigma_up            = std::min(sigma_up, sigma_to);
+                    float sigma_down_sq = sigma_to_sq - sigma_up * sigma_up;
+                    sigma_down          = sigma_down_sq > 0.0f ? std::sqrt(sigma_down_sq) : 0.0f;
+                }
+
+                if (sigma_down == 0.0f || !have_old_sigma) {
+                    float dt            = sigma_down - sigma_from;
+                    float* vec_x        = (float*)x->data;
+                    float* vec_denoised = (float*)denoised->data;
+
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        float d  = (vec_x[j] - vec_denoised[j]) / sigma_from;
+                        vec_x[j] = vec_x[j] + d * dt;
+                    }
+                } else {
+                    float t      = t_fn(sigma_from);
+                    float t_old  = t_fn(old_sigma_down);
+                    float t_next = t_fn(sigma_down);
+                    float t_prev = t_fn(sigmas[i - 1]);
+                    float h      = t_next - t;
+                    float c2     = (t_prev - t_old) / h;
+
+                    float phi1_val = phi1_fn(-h);
+                    float phi2_val = phi2_fn(-h);
+                    float b1       = phi1_val - phi2_val / c2;
+                    float b2       = phi2_val / c2;
+
+                    if (!std::isfinite(b1)) {
+                        b1 = 0.0f;
+                    }
+                    if (!std::isfinite(b2)) {
+                        b2 = 0.0f;
+                    }
+
+                    float sigma_h           = sigma_fn(h);
+                    float* vec_x            = (float*)x->data;
+                    float* vec_denoised     = (float*)denoised->data;
+                    float* vec_old_denoised = (float*)old_denoised->data;
+
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] = sigma_h * vec_x[j] + h * (b1 * vec_denoised[j] + b2 * vec_old_denoised[j]);
+                    }
+                }
+
+                if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
+                    ggml_ext_im_set_randn_f32(noise, rng);
+                    float* vec_x     = (float*)x->data;
+                    float* vec_noise = (float*)noise->data;
+
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] = vec_x[j] + vec_noise[j] * sigma_up;
+                    }
+                }
+
+                float* vec_old_denoised = (float*)old_denoised->data;
+                float* vec_denoised     = (float*)denoised->data;
+                for (int j = 0; j < ggml_nelements(x); j++) {
+                    vec_old_denoised[j] = vec_denoised[j];
+                }
+
+                old_sigma_down = sigma_down;
+                have_old_sigma = true;
+            }
+        } break;
+        case RES_2S_SAMPLE_METHOD:  // Res 2s sampler
+        {
+            struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
+            struct ggml_tensor* x0    = ggml_dup_tensor(work_ctx, x);
+            struct ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
+
+            const float c2 = 0.5f;
+            auto t_fn      = [](float sigma) -> float { return -logf(sigma); };
+            auto phi1_fn   = [](float t) -> float {
+                if (fabsf(t) < 1e-6f) {
+                    return 1.0f + t * 0.5f + (t * t) / 6.0f;
+                }
+                return (expf(t) - 1.0f) / t;
+            };
+            auto phi2_fn = [&](float t) -> float {
+                if (fabsf(t) < 1e-6f) {
+                    return 0.5f + t / 6.0f + (t * t) / 24.0f;
+                }
+                float phi1_val = phi1_fn(t);
+                return (phi1_val - 1.0f) / t;
+            };
+
+            for (int i = 0; i < steps; i++) {
+                float sigma_from = sigmas[i];
+                float sigma_to   = sigmas[i + 1];
+
+                ggml_tensor* denoised = model(x, sigma_from, -(i + 1));
+                if (denoised == nullptr) {
+                    return false;
+                }
+
+                float sigma_up   = 0.0f;
+                float sigma_down = sigma_to;
+                if (eta > 0.0f) {
+                    float sigma_from_sq = sigma_from * sigma_from;
+                    float sigma_to_sq   = sigma_to * sigma_to;
+                    if (sigma_from_sq > 0.0f) {
+                        float term = sigma_to_sq * (sigma_from_sq - sigma_to_sq) / sigma_from_sq;
+                        if (term > 0.0f) {
+                            sigma_up = eta * std::sqrt(term);
+                        }
+                    }
+                    sigma_up            = std::min(sigma_up, sigma_to);
+                    float sigma_down_sq = sigma_to_sq - sigma_up * sigma_up;
+                    sigma_down          = sigma_down_sq > 0.0f ? std::sqrt(sigma_down_sq) : 0.0f;
+                }
+
+                float* vec_x  = (float*)x->data;
+                float* vec_x0 = (float*)x0->data;
+                for (int j = 0; j < ggml_nelements(x); j++) {
+                    vec_x0[j] = vec_x[j];
+                }
+
+                if (sigma_down == 0.0f || sigma_from == 0.0f) {
+                    float* vec_denoised = (float*)denoised->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] = vec_denoised[j];
+                    }
+                } else {
+                    float t      = t_fn(sigma_from);
+                    float t_next = t_fn(sigma_down);
+                    float h      = t_next - t;
+
+                    float a21      = c2 * phi1_fn(-h * c2);
+                    float phi1_val = phi1_fn(-h);
+                    float phi2_val = phi2_fn(-h);
+                    float b2       = phi2_val / c2;
+                    float b1       = phi1_val - b2;
+
+                    float sigma_c2 = expf(-(t + h * c2));
+
+                    float* vec_denoised = (float*)denoised->data;
+                    float* vec_x2       = (float*)x2->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        float eps1 = vec_denoised[j] - vec_x0[j];
+                        vec_x2[j]  = vec_x0[j] + h * a21 * eps1;
+                    }
+
+                    ggml_tensor* denoised2 = model(x2, sigma_c2, i + 1);
+                    if (denoised2 == nullptr) {
+                        return false;
+                    }
+                    float* vec_denoised2 = (float*)denoised2->data;
+
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        float eps1 = vec_denoised[j] - vec_x0[j];
+                        float eps2 = vec_denoised2[j] - vec_x0[j];
+                        vec_x[j]   = vec_x0[j] + h * (b1 * eps1 + b2 * eps2);
+                    }
+                }
+
+                if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
+                    ggml_ext_im_set_randn_f32(noise, rng);
+                    float* vec_x     = (float*)x->data;
+                    float* vec_noise = (float*)noise->data;
+
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] = vec_x[j] + vec_noise[j] * sigma_up;
+                    }
+                }
+            }
+        } break;
 
         default:
             LOG_ERROR("Attempting to sample with nonexisting sample method %i", method);

diffusion_model.hpp

@@ -38,7 +38,7 @@ struct DiffusionModel {
     virtual size_t get_params_buffer_size()                                             = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
     virtual int64_t get_adm_in_channels()                            = 0;
-    virtual void set_flash_attn_enabled(bool enabled)                = 0;
+    virtual void set_flash_attention_enabled(bool enabled)           = 0;
     virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };
 
@@ -84,7 +84,7 @@ struct UNetModel : public DiffusionModel {
         return unet.unet.adm_in_channels;
     }
 
-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
         unet.set_flash_attention_enabled(enabled);
     }
 
@@ -149,7 +149,7 @@ struct MMDiTModel : public DiffusionModel {
         return 768 + 1280;
     }
 
-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
         mmdit.set_flash_attention_enabled(enabled);
     }
 
@@ -215,7 +215,7 @@ struct FluxModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
         flux.set_flash_attention_enabled(enabled);
     }
 
@@ -286,7 +286,7 @@ struct WanModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
         wan.set_flash_attention_enabled(enabled);
     }
 
@@ -357,7 +357,7 @@ struct QwenImageModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
         qwen_image.set_flash_attention_enabled(enabled);
     }
 
@@ -424,7 +424,7 @@ struct ZImageModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
         z_image.set_flash_attention_enabled(enabled);
     }
 

docs/z_image.md

@@ -7,6 +7,9 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 - Download Z-Image-Turbo
     - safetensors: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/diffusion_models
     - gguf: https://huggingface.co/leejet/Z-Image-Turbo-GGUF/tree/main
+- Download Z-Image
+    - safetensors: https://huggingface.co/Comfy-Org/z_image/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/unsloth/Z-Image-GGUF/tree/main
 - Download vae
     - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
 - Download Qwen3 4b
@@ -15,12 +18,22 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 
 ## Examples
 
+### Z-Image-Turbo
+
 ```
 .\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
 ```
 
 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
 
+### Z-Image-Base
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\z_image_bf16.safetensors --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
+```
+
+<img width="256" alt="z-image example" src="../assets/z_image/base_bf16.png" />
+
 ## Comparison of Different Quantization Types
 
 | bf16 | q8_0 | q6_K | q5_0 | q4_K | q4_0 | q3_K | q2_K|

examples/cli/README.md

@@ -52,7 +52,8 @@ Context Options:
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
   --vae-on-cpu                             keep vae in cpu (for low vram)
-  --diffusion-fa                           use flash attention in the diffusion model
+  --fa                                     use flash attention
+  --diffusion-fa                           use flash attention in the diffusion model only
   --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
   --vae-conv-direct                        use ggml_conv2d_direct in the vae model
   --circular                               enable circular padding for convolutions
@@ -107,14 +108,14 @@ Generation Options:
                                            medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
+  --eta <float>                            eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
   --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
   --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM and TCD (default: 0)
+  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
   --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
@@ -123,12 +124,12 @@ Generation Options:
   --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
-  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
-  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd,
+                                           res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
+                                           tcd, res_multistep, res_2s] default: euler for Flux/SD3/Wan, euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm], default: discrete
+                                           kl_optimal, lcm, bong_tangent], default: discrete
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])

examples/common/common.hpp

@@ -457,6 +457,7 @@ struct SDContextParams {
     bool control_net_cpu        = false;
     bool clip_on_cpu            = false;
     bool vae_on_cpu             = false;
+    bool flash_attn             = false;
     bool diffusion_flash_attn   = false;
     bool diffusion_conv_direct  = false;
     bool vae_conv_direct        = false;
@@ -615,9 +616,13 @@ struct SDContextParams {
              "--vae-on-cpu",
              "keep vae in cpu (for low vram)",
              true, &vae_on_cpu},
+            {"",
+             "--fa",
+             "use flash attention",
+             true, &flash_attn},
             {"",
              "--diffusion-fa",
-             "use flash attention in the diffusion model",
+             "use flash attention in the diffusion model only",
              true, &diffusion_flash_attn},
             {"",
              "--diffusion-conv-direct",
@@ -904,6 +909,7 @@ struct SDContextParams {
             << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
             << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
             << "  vae_on_cpu: " << (vae_on_cpu ? "true" : "false") << ",\n"
+            << "  flash_attn: " << (flash_attn ? "true" : "false") << ",\n"
             << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
             << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
             << "  vae_conv_direct: " << (vae_conv_direct ? "true" : "false") << ",\n"
@@ -968,6 +974,7 @@ struct SDContextParams {
             clip_on_cpu,
             control_net_cpu,
             vae_on_cpu,
+            flash_attn,
             diffusion_flash_attn,
             taesd_preview,
             diffusion_conv_direct,
@@ -1478,17 +1485,17 @@ struct SDGenerationParams {
              on_seed_arg},
             {"",
              "--sampling-method",
-             "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd] "
+             "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s] "
              "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
              on_sample_method_arg},
             {"",
              "--high-noise-sampling-method",
-             "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd]"
+             "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s]"
              " default: euler for Flux/SD3/Wan, euler_a otherwise",
              on_high_noise_sample_method_arg},
             {"",
              "--scheduler",
-             "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm], default: discrete",
+             "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default: discrete",
              on_scheduler_arg},
             {"",
              "--sigmas",

examples/server/README.md

@@ -44,7 +44,8 @@ Context Options:
   --clip-on-cpu                            keep clip in cpu (for low vram)
   --vae-on-cpu                             keep vae in cpu (for low vram)
   --mmap                                   whether to memory-map model
-  --diffusion-fa                           use flash attention in the diffusion model
+  --fa                                     use flash attention
+  --diffusion-fa                           use flash attention in the diffusion model only
   --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
   --vae-conv-direct                        use ggml_conv2d_direct in the vae model
   --circular                               enable circular padding for convolutions
@@ -99,14 +100,14 @@ Default Generation Options:
                                            medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
+  --eta <float>                            eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
   --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
   --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM and TCD (default: 0)
+  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
   --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
@@ -115,12 +116,12 @@ Default Generation Options:
   --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
-  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
-  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd,
+                                           res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
+                                           tcd, res_multistep, res_2s] default: euler for Flux/SD3/Wan, euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm], default: discrete
+                                           kl_optimal, lcm, bong_tangent], default: discrete
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])

examples/server/main.cpp

@@ -785,7 +785,11 @@ int main(int argc, const char** argv) {
                     {"lcm", LCM_SAMPLE_METHOD},
                     {"ddim", DDIM_TRAILING_SAMPLE_METHOD},
                     {"dpm++ 2m", DPMPP2M_SAMPLE_METHOD},
-                    {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD}};
+                    {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD},
+                    {"res multistep", RES_MULTISTEP_SAMPLE_METHOD},
+                    {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
+                    {"res 2s", RES_2S_SAMPLE_METHOD},
+                    {"k_res_2s", RES_2S_SAMPLE_METHOD}};
                 auto it            = hardcoded.find(name);
                 if (it != hardcoded.end()) return it->second;
                 return SAMPLE_METHOD_COUNT;

ggml

@@ -1 +1 @@
-Subproject commit 8891ab6fc742ac1198736d3da3b73c730e42af84
+Subproject commit a8db410a252c8c8f2d120c6f2e7133ebe032f35d

ggml_extend.hpp

@@ -2623,7 +2623,7 @@ public:
             v = v_proj->forward(ctx, x);
         }
 
-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask);  // [N, n_token, embed_dim]
+        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, false);  // [N, n_token, embed_dim]
 
         x = out_proj->forward(ctx, x);  // [N, n_token, embed_dim]
         return x;

name_conversion.cpp

@@ -842,6 +842,7 @@ std::string convert_sep_to_dot(std::string name) {
         "conv_in",
         "conv_out",
         "lora_down",
+        "lora_mid",
         "lora_up",
         "diff_b",
         "hada_w1_a",
@@ -997,10 +998,13 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
     if (is_lora) {
         std::map<std::string, std::string> lora_suffix_map = {
             {".lora_down.weight", ".weight.lora_down"},
+            {".lora_mid.weight", ".weight.lora_mid"},
             {".lora_up.weight", ".weight.lora_up"},
             {".lora.down.weight", ".weight.lora_down"},
+            {".lora.mid.weight", ".weight.lora_mid"},
             {".lora.up.weight", ".weight.lora_up"},
             {"_lora.down.weight", ".weight.lora_down"},
+            {"_lora.mid.weight", ".weight.lora_mid"},
             {"_lora.up.weight", ".weight.lora_up"},
             {".lora_A.weight", ".weight.lora_down"},
             {".lora_B.weight", ".weight.lora_up"},

qwen_image.hpp

@@ -212,7 +212,7 @@ namespace Qwen {
 
             blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
             blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
-            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU));
+            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU, true));
 
             blocks["attn"] = std::shared_ptr<GGMLBlock>(new QwenImageAttention(dim,
                                                                                attention_head_dim,

stable-diffusion.cpp

@@ -67,6 +67,8 @@ const char* sampling_methods_str[] = {
     "LCM",
     "DDIM \"trailing\"",
     "TCD",
+    "Res Multistep",
+    "Res 2s",
 };
 
 /*================================================== Helper Functions ================================================*/
@@ -443,7 +445,7 @@ public:
                     }
                 }
                 if (is_chroma) {
-                    if (sd_ctx_params->diffusion_flash_attn && sd_ctx_params->chroma_use_dit_mask) {
+                    if ((sd_ctx_params->flash_attn || 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. "
@@ -569,14 +571,6 @@ public:
                 }
             }
 
-            if (sd_ctx_params->diffusion_flash_attn) {
-                LOG_INFO("Using flash attention in the diffusion model");
-                diffusion_model->set_flash_attn_enabled(true);
-                if (high_noise_diffusion_model) {
-                    high_noise_diffusion_model->set_flash_attn_enabled(true);
-                }
-            }
-
             cond_stage_model->alloc_params_buffer();
             cond_stage_model->get_param_tensors(tensors);
 
@@ -723,6 +717,28 @@ public:
                 pmid_model->get_param_tensors(tensors, "pmid");
             }
 
+            if (sd_ctx_params->flash_attn) {
+                LOG_INFO("Using flash attention");
+                cond_stage_model->set_flash_attention_enabled(true);
+                if (clip_vision) {
+                    clip_vision->set_flash_attention_enabled(true);
+                }
+                if (first_stage_model) {
+                    first_stage_model->set_flash_attention_enabled(true);
+                }
+                if (tae_first_stage) {
+                    tae_first_stage->set_flash_attention_enabled(true);
+                }
+            }
+
+            if (sd_ctx_params->flash_attn || sd_ctx_params->diffusion_flash_attn) {
+                LOG_INFO("Using flash attention in the diffusion model");
+                diffusion_model->set_flash_attention_enabled(true);
+                if (high_noise_diffusion_model) {
+                    high_noise_diffusion_model->set_flash_attention_enabled(true);
+                }
+            }
+
             diffusion_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
             if (high_noise_diffusion_model) {
                 high_noise_diffusion_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
@@ -2743,6 +2759,8 @@ const char* sample_method_to_str[] = {
     "lcm",
     "ddim_trailing",
     "tcd",
+    "res_multistep",
+    "res_2s",
 };
 
 const char* sd_sample_method_name(enum sample_method_t sample_method) {
@@ -2772,6 +2790,7 @@ const char* scheduler_to_str[] = {
     "smoothstep",
     "kl_optimal",
     "lcm",
+    "bong_tangent",
 };
 
 const char* sd_scheduler_name(enum scheduler_t scheduler) {
@@ -2937,6 +2956,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "keep_clip_on_cpu: %s\n"
              "keep_control_net_on_cpu: %s\n"
              "keep_vae_on_cpu: %s\n"
+             "flash_attn: %s\n"
              "diffusion_flash_attn: %s\n"
              "circular_x: %s\n"
              "circular_y: %s\n"
@@ -2968,6 +2988,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              BOOL_STR(sd_ctx_params->keep_clip_on_cpu),
              BOOL_STR(sd_ctx_params->keep_control_net_on_cpu),
              BOOL_STR(sd_ctx_params->keep_vae_on_cpu),
+             BOOL_STR(sd_ctx_params->flash_attn),
              BOOL_STR(sd_ctx_params->diffusion_flash_attn),
              BOOL_STR(sd_ctx_params->circular_x),
              BOOL_STR(sd_ctx_params->circular_y),

stable-diffusion.h

@@ -48,6 +48,8 @@ enum sample_method_t {
     LCM_SAMPLE_METHOD,
     DDIM_TRAILING_SAMPLE_METHOD,
     TCD_SAMPLE_METHOD,
+    RES_MULTISTEP_SAMPLE_METHOD,
+    RES_2S_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -62,6 +64,7 @@ enum scheduler_t {
     SMOOTHSTEP_SCHEDULER,
     KL_OPTIMAL_SCHEDULER,
     LCM_SCHEDULER,
+    BONG_TANGENT_SCHEDULER,
     SCHEDULER_COUNT
 };
 
@@ -186,6 +189,7 @@ typedef struct {
     bool keep_clip_on_cpu;
     bool keep_control_net_on_cpu;
     bool keep_vae_on_cpu;
+    bool flash_attn;
     bool diffusion_flash_attn;
     bool tae_preview_only;
     bool diffusion_conv_direct;

vae.hpp

@@ -141,7 +141,7 @@ public:
             v = ggml_reshape_3d(ctx->ggml_ctx, v, c, h * w, n);                        // [N, h * w, in_channels]
         }
 
-        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, true, false);
+        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, true, ctx->flash_attn_enabled);
 
         if (use_linear) {
             h_ = proj_out->forward(ctx, h_);  // [N, h * w, in_channels]

wan.hpp

@@ -573,7 +573,7 @@ namespace WAN {
             v      = ggml_reshape_3d(ctx->ggml_ctx, v, h * w, c, n);  // [t, c, h * w]
 
             v = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, v, 1, 0, 2, 3));                           // [t, h * w, c]
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, true, false);  // [t, h * w, c]
+            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, true, ctx->flash_attn_enabled);  // [t, h * w, c]
 
             x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [t, c, h * w]
             x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, c, n);                             // [t, c, h, w]