feat: support random seed flag (#1163)

.github/workflows/build.yml

@@ -163,7 +163,7 @@ jobs:
           - build: "avx512"
             defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
           - build: "cuda12"
-            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120'"
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120' -DCMAKE_CUDA_FLAGS='-Xcudafe \"--diag_suppress=177\" -Xcudafe \"--diag_suppress=550\"'"
           - build: 'vulkan'
             defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
     steps:
@@ -191,13 +191,17 @@ jobs:
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
 
+      - name: Activate MSVC environment
+        id: msvc_dev_cmd
+        uses: ilammy/msvc-dev-cmd@v1
+
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release
+          cmake .. -DCMAKE_CXX_FLAGS='/bigobj' -G Ninja -DCMAKE_C_COMPILER=cl.exe -DCMAKE_CXX_COMPILER=cl.exe ${{ matrix.defines }}
+          cmake --build .
 
       - name: Check AVX512F support
         id: check_avx512f

README.md

@@ -52,7 +52,7 @@ API and command-line option may change frequently.***
     - [Ovis-Image](./docs/ovis_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
-    - [Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
+    - [Qwen Image Edit series](./docs/qwen_image_edit.md)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
   - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
@@ -132,7 +132,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [FLUX.1-Kontext-dev](./docs/kontext.md)
 - [Chroma](./docs/chroma.md)
 - [🔥Qwen Image](./docs/qwen_image.md)
-- [🔥Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
+- [🔥Qwen Image Edit series](./docs/qwen_image_edit.md)
 - [🔥Wan2.1/Wan2.2](./docs/wan.md)
 - [🔥Z-Image](./docs/z_image.md)
 - [Ovis-Image](./docs/ovis_image.md)
@@ -143,6 +143,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [Using TAESD to faster decoding](./docs/taesd.md)
 - [Docker](./docs/docker.md)
 - [Quantization and GGUF](./docs/quantization_and_gguf.md)
+- [Inference acceleration via caching](./docs/caching.md)
 
 ## Bindings
 

cache_dit.hpp

@@ -0,0 +1,975 @@
+#ifndef __CACHE_DIT_HPP__
+#define __CACHE_DIT_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <limits>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "ggml_extend.hpp"
+
+struct DBCacheConfig {
+    bool enabled                        = false;
+    int Fn_compute_blocks               = 8;
+    int Bn_compute_blocks               = 0;
+    float residual_diff_threshold       = 0.08f;
+    int max_warmup_steps                = 8;
+    int max_cached_steps                = -1;
+    int max_continuous_cached_steps     = -1;
+    float max_accumulated_residual_diff = -1.0f;
+    std::vector<int> steps_computation_mask;
+    bool scm_policy_dynamic = true;
+};
+
+struct TaylorSeerConfig {
+    bool enabled            = false;
+    int n_derivatives       = 1;
+    int max_warmup_steps    = 2;
+    int skip_interval_steps = 1;
+};
+
+struct CacheDitConfig {
+    DBCacheConfig dbcache;
+    TaylorSeerConfig taylorseer;
+    int double_Fn_blocks = -1;
+    int double_Bn_blocks = -1;
+    int single_Fn_blocks = -1;
+    int single_Bn_blocks = -1;
+};
+
+struct TaylorSeerState {
+    int n_derivatives      = 1;
+    int current_step       = -1;
+    int last_computed_step = -1;
+    std::vector<std::vector<float>> dY_prev;
+    std::vector<std::vector<float>> dY_current;
+
+    void init(int n_deriv, size_t hidden_size) {
+        n_derivatives = n_deriv;
+        int order     = n_derivatives + 1;
+        dY_prev.resize(order);
+        dY_current.resize(order);
+        for (int i = 0; i < order; i++) {
+            dY_prev[i].clear();
+            dY_current[i].clear();
+        }
+        current_step       = -1;
+        last_computed_step = -1;
+    }
+
+    void reset() {
+        for (auto& v : dY_prev)
+            v.clear();
+        for (auto& v : dY_current)
+            v.clear();
+        current_step       = -1;
+        last_computed_step = -1;
+    }
+
+    bool can_approximate() const {
+        return last_computed_step >= n_derivatives && !dY_prev.empty() && !dY_prev[0].empty();
+    }
+
+    void update_derivatives(const float* Y, size_t size, int step) {
+        int order = n_derivatives + 1;
+        dY_prev   = dY_current;
+        dY_current[0].resize(size);
+        for (size_t i = 0; i < size; i++) {
+            dY_current[0][i] = Y[i];
+        }
+
+        int window = step - last_computed_step;
+        if (window <= 0)
+            window = 1;
+
+        for (int d = 0; d < n_derivatives; d++) {
+            if (!dY_prev[d].empty() && dY_prev[d].size() == size) {
+                dY_current[d + 1].resize(size);
+                for (size_t i = 0; i < size; i++) {
+                    dY_current[d + 1][i] = (dY_current[d][i] - dY_prev[d][i]) / static_cast<float>(window);
+                }
+            } else {
+                dY_current[d + 1].clear();
+            }
+        }
+
+        current_step       = step;
+        last_computed_step = step;
+    }
+
+    void approximate(float* output, size_t size, int target_step) const {
+        if (!can_approximate() || dY_prev[0].size() != size) {
+            return;
+        }
+
+        int elapsed = target_step - last_computed_step;
+        if (elapsed <= 0)
+            elapsed = 1;
+
+        std::fill(output, output + size, 0.0f);
+        float factorial = 1.0f;
+        int order       = static_cast<int>(dY_prev.size());
+
+        for (int o = 0; o < order; o++) {
+            if (dY_prev[o].empty() || dY_prev[o].size() != size)
+                continue;
+            if (o > 0)
+                factorial *= static_cast<float>(o);
+            float coeff = std::pow(static_cast<float>(elapsed), o) / factorial;
+            for (size_t i = 0; i < size; i++) {
+                output[i] += coeff * dY_prev[o][i];
+            }
+        }
+    }
+};
+
+struct BlockCacheEntry {
+    std::vector<float> residual_img;
+    std::vector<float> residual_txt;
+    std::vector<float> residual;
+    std::vector<float> prev_img;
+    std::vector<float> prev_txt;
+    std::vector<float> prev_output;
+    bool has_prev = false;
+};
+
+struct CacheDitState {
+    CacheDitConfig config;
+    bool initialized = false;
+
+    int total_double_blocks = 0;
+    int total_single_blocks = 0;
+    size_t hidden_size      = 0;
+
+    int current_step     = -1;
+    int total_steps      = 0;
+    int warmup_remaining = 0;
+    std::vector<int> cached_steps;
+    int continuous_cached_steps     = 0;
+    float accumulated_residual_diff = 0.0f;
+
+    std::vector<BlockCacheEntry> double_block_cache;
+    std::vector<BlockCacheEntry> single_block_cache;
+
+    std::vector<float> Fn_residual_img;
+    std::vector<float> Fn_residual_txt;
+    std::vector<float> prev_Fn_residual_img;
+    std::vector<float> prev_Fn_residual_txt;
+    bool has_prev_Fn_residual = false;
+
+    std::vector<float> Bn_buffer_img;
+    std::vector<float> Bn_buffer_txt;
+    std::vector<float> Bn_buffer;
+    bool has_Bn_buffer = false;
+
+    TaylorSeerState taylor_state;
+
+    bool can_cache_this_step  = false;
+    bool is_caching_this_step = false;
+
+    int total_blocks_computed = 0;
+    int total_blocks_cached   = 0;
+
+    void init(const CacheDitConfig& cfg, int num_double_blocks, int num_single_blocks, size_t h_size) {
+        config              = cfg;
+        total_double_blocks = num_double_blocks;
+        total_single_blocks = num_single_blocks;
+        hidden_size         = h_size;
+
+        initialized = cfg.dbcache.enabled || cfg.taylorseer.enabled;
+
+        if (!initialized)
+            return;
+
+        warmup_remaining = cfg.dbcache.max_warmup_steps;
+        double_block_cache.resize(total_double_blocks);
+        single_block_cache.resize(total_single_blocks);
+
+        if (cfg.taylorseer.enabled) {
+            taylor_state.init(cfg.taylorseer.n_derivatives, h_size);
+        }
+
+        reset_runtime();
+    }
+
+    void reset_runtime() {
+        current_step     = -1;
+        total_steps      = 0;
+        warmup_remaining = config.dbcache.max_warmup_steps;
+        cached_steps.clear();
+        continuous_cached_steps   = 0;
+        accumulated_residual_diff = 0.0f;
+
+        for (auto& entry : double_block_cache) {
+            entry.residual_img.clear();
+            entry.residual_txt.clear();
+            entry.prev_img.clear();
+            entry.prev_txt.clear();
+            entry.has_prev = false;
+        }
+
+        for (auto& entry : single_block_cache) {
+            entry.residual.clear();
+            entry.prev_output.clear();
+            entry.has_prev = false;
+        }
+
+        Fn_residual_img.clear();
+        Fn_residual_txt.clear();
+        prev_Fn_residual_img.clear();
+        prev_Fn_residual_txt.clear();
+        has_prev_Fn_residual = false;
+
+        Bn_buffer_img.clear();
+        Bn_buffer_txt.clear();
+        Bn_buffer.clear();
+        has_Bn_buffer = false;
+
+        taylor_state.reset();
+
+        can_cache_this_step  = false;
+        is_caching_this_step = false;
+
+        total_blocks_computed = 0;
+        total_blocks_cached   = 0;
+    }
+
+    bool enabled() const {
+        return initialized && (config.dbcache.enabled || config.taylorseer.enabled);
+    }
+
+    void begin_step(int step_index, float sigma = 0.0f) {
+        if (!enabled())
+            return;
+        if (step_index == current_step)
+            return;
+
+        current_step = step_index;
+        total_steps++;
+
+        bool in_warmup = warmup_remaining > 0;
+        if (in_warmup) {
+            warmup_remaining--;
+        }
+
+        bool scm_allows_cache = true;
+        if (!config.dbcache.steps_computation_mask.empty()) {
+            if (step_index < static_cast<int>(config.dbcache.steps_computation_mask.size())) {
+                scm_allows_cache = (config.dbcache.steps_computation_mask[step_index] == 0);
+                if (!config.dbcache.scm_policy_dynamic && scm_allows_cache) {
+                    can_cache_this_step  = true;
+                    is_caching_this_step = false;
+                    return;
+                }
+            }
+        }
+
+        bool max_cached_ok = (config.dbcache.max_cached_steps < 0) ||
+                             (static_cast<int>(cached_steps.size()) < config.dbcache.max_cached_steps);
+
+        bool max_cont_ok = (config.dbcache.max_continuous_cached_steps < 0) ||
+                           (continuous_cached_steps < config.dbcache.max_continuous_cached_steps);
+
+        bool accum_ok = (config.dbcache.max_accumulated_residual_diff < 0.0f) ||
+                        (accumulated_residual_diff < config.dbcache.max_accumulated_residual_diff);
+
+        can_cache_this_step  = !in_warmup && scm_allows_cache && max_cached_ok && max_cont_ok && accum_ok && has_prev_Fn_residual;
+        is_caching_this_step = false;
+    }
+
+    void end_step(bool was_cached) {
+        if (was_cached) {
+            cached_steps.push_back(current_step);
+            continuous_cached_steps++;
+        } else {
+            continuous_cached_steps = 0;
+        }
+    }
+
+    static float calculate_residual_diff(const float* prev, const float* curr, size_t size) {
+        if (size == 0)
+            return 0.0f;
+
+        float sum_diff = 0.0f;
+        float sum_abs  = 0.0f;
+
+        for (size_t i = 0; i < size; i++) {
+            sum_diff += std::fabs(prev[i] - curr[i]);
+            sum_abs += std::fabs(prev[i]);
+        }
+
+        return sum_diff / (sum_abs + 1e-6f);
+    }
+
+    static float calculate_residual_diff(const std::vector<float>& prev, const std::vector<float>& curr) {
+        if (prev.size() != curr.size() || prev.empty())
+            return 1.0f;
+        return calculate_residual_diff(prev.data(), curr.data(), prev.size());
+    }
+
+    int get_double_Fn_blocks() const {
+        return (config.double_Fn_blocks >= 0) ? config.double_Fn_blocks : config.dbcache.Fn_compute_blocks;
+    }
+
+    int get_double_Bn_blocks() const {
+        return (config.double_Bn_blocks >= 0) ? config.double_Bn_blocks : config.dbcache.Bn_compute_blocks;
+    }
+
+    int get_single_Fn_blocks() const {
+        return (config.single_Fn_blocks >= 0) ? config.single_Fn_blocks : config.dbcache.Fn_compute_blocks;
+    }
+
+    int get_single_Bn_blocks() const {
+        return (config.single_Bn_blocks >= 0) ? config.single_Bn_blocks : config.dbcache.Bn_compute_blocks;
+    }
+
+    bool is_Fn_double_block(int block_idx) const {
+        return block_idx < get_double_Fn_blocks();
+    }
+
+    bool is_Bn_double_block(int block_idx) const {
+        int Bn = get_double_Bn_blocks();
+        return Bn > 0 && block_idx >= (total_double_blocks - Bn);
+    }
+
+    bool is_Mn_double_block(int block_idx) const {
+        return !is_Fn_double_block(block_idx) && !is_Bn_double_block(block_idx);
+    }
+
+    bool is_Fn_single_block(int block_idx) const {
+        return block_idx < get_single_Fn_blocks();
+    }
+
+    bool is_Bn_single_block(int block_idx) const {
+        int Bn = get_single_Bn_blocks();
+        return Bn > 0 && block_idx >= (total_single_blocks - Bn);
+    }
+
+    bool is_Mn_single_block(int block_idx) const {
+        return !is_Fn_single_block(block_idx) && !is_Bn_single_block(block_idx);
+    }
+
+    void store_Fn_residual(const float* img, const float* txt, size_t img_size, size_t txt_size, const float* input_img, const float* input_txt) {
+        Fn_residual_img.resize(img_size);
+        Fn_residual_txt.resize(txt_size);
+
+        for (size_t i = 0; i < img_size; i++) {
+            Fn_residual_img[i] = img[i] - input_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            Fn_residual_txt[i] = txt[i] - input_txt[i];
+        }
+    }
+
+    bool check_cache_decision() {
+        if (!can_cache_this_step) {
+            is_caching_this_step = false;
+            return false;
+        }
+
+        if (!has_prev_Fn_residual || prev_Fn_residual_img.empty()) {
+            is_caching_this_step = false;
+            return false;
+        }
+
+        float diff_img = calculate_residual_diff(prev_Fn_residual_img, Fn_residual_img);
+        float diff_txt = calculate_residual_diff(prev_Fn_residual_txt, Fn_residual_txt);
+        float diff     = (diff_img + diff_txt) / 2.0f;
+
+        if (diff < config.dbcache.residual_diff_threshold) {
+            is_caching_this_step = true;
+            accumulated_residual_diff += diff;
+            return true;
+        }
+
+        is_caching_this_step = false;
+        return false;
+    }
+
+    void update_prev_Fn_residual() {
+        prev_Fn_residual_img = Fn_residual_img;
+        prev_Fn_residual_txt = Fn_residual_txt;
+        has_prev_Fn_residual = !prev_Fn_residual_img.empty();
+    }
+
+    void store_double_block_residual(int block_idx, const float* img, const float* txt, size_t img_size, size_t txt_size, const float* prev_img, const float* prev_txt) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(double_block_cache.size()))
+            return;
+
+        BlockCacheEntry& entry = double_block_cache[block_idx];
+
+        entry.residual_img.resize(img_size);
+        entry.residual_txt.resize(txt_size);
+        for (size_t i = 0; i < img_size; i++) {
+            entry.residual_img[i] = img[i] - prev_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            entry.residual_txt[i] = txt[i] - prev_txt[i];
+        }
+
+        entry.prev_img.resize(img_size);
+        entry.prev_txt.resize(txt_size);
+        for (size_t i = 0; i < img_size; i++) {
+            entry.prev_img[i] = img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            entry.prev_txt[i] = txt[i];
+        }
+        entry.has_prev = true;
+    }
+
+    void apply_double_block_cache(int block_idx, float* img, float* txt, size_t img_size, size_t txt_size) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(double_block_cache.size()))
+            return;
+
+        const BlockCacheEntry& entry = double_block_cache[block_idx];
+        if (entry.residual_img.size() != img_size || entry.residual_txt.size() != txt_size)
+            return;
+
+        for (size_t i = 0; i < img_size; i++) {
+            img[i] += entry.residual_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            txt[i] += entry.residual_txt[i];
+        }
+
+        total_blocks_cached++;
+    }
+
+    void store_single_block_residual(int block_idx, const float* output, size_t size, const float* input) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(single_block_cache.size()))
+            return;
+
+        BlockCacheEntry& entry = single_block_cache[block_idx];
+
+        entry.residual.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.residual[i] = output[i] - input[i];
+        }
+
+        entry.prev_output.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.prev_output[i] = output[i];
+        }
+        entry.has_prev = true;
+    }
+
+    void apply_single_block_cache(int block_idx, float* output, size_t size) {
+        if (block_idx < 0 || block_idx >= static_cast<int>(single_block_cache.size()))
+            return;
+
+        const BlockCacheEntry& entry = single_block_cache[block_idx];
+        if (entry.residual.size() != size)
+            return;
+
+        for (size_t i = 0; i < size; i++) {
+            output[i] += entry.residual[i];
+        }
+
+        total_blocks_cached++;
+    }
+
+    void store_Bn_buffer(const float* img, const float* txt, size_t img_size, size_t txt_size, const float* Bn_start_img, const float* Bn_start_txt) {
+        Bn_buffer_img.resize(img_size);
+        Bn_buffer_txt.resize(txt_size);
+
+        for (size_t i = 0; i < img_size; i++) {
+            Bn_buffer_img[i] = img[i] - Bn_start_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            Bn_buffer_txt[i] = txt[i] - Bn_start_txt[i];
+        }
+        has_Bn_buffer = true;
+    }
+
+    void apply_Bn_buffer(float* img, float* txt, size_t img_size, size_t txt_size) {
+        if (!has_Bn_buffer)
+            return;
+        if (Bn_buffer_img.size() != img_size || Bn_buffer_txt.size() != txt_size)
+            return;
+
+        for (size_t i = 0; i < img_size; i++) {
+            img[i] += Bn_buffer_img[i];
+        }
+        for (size_t i = 0; i < txt_size; i++) {
+            txt[i] += Bn_buffer_txt[i];
+        }
+    }
+
+    void taylor_update(const float* hidden_state, size_t size) {
+        if (!config.taylorseer.enabled)
+            return;
+        taylor_state.update_derivatives(hidden_state, size, current_step);
+    }
+
+    bool taylor_can_approximate() const {
+        return config.taylorseer.enabled && taylor_state.can_approximate();
+    }
+
+    void taylor_approximate(float* output, size_t size) {
+        if (!config.taylorseer.enabled)
+            return;
+        taylor_state.approximate(output, size, current_step);
+    }
+
+    bool should_use_taylor_this_step() const {
+        if (!config.taylorseer.enabled)
+            return false;
+        if (current_step < config.taylorseer.max_warmup_steps)
+            return false;
+
+        int interval = config.taylorseer.skip_interval_steps;
+        if (interval <= 0)
+            interval = 1;
+
+        return (current_step % (interval + 1)) != 0;
+    }
+
+    void log_metrics() const {
+        if (!enabled())
+            return;
+
+        int total_blocks  = total_blocks_computed + total_blocks_cached;
+        float cache_ratio = (total_blocks > 0) ? (static_cast<float>(total_blocks_cached) / total_blocks * 100.0f) : 0.0f;
+
+        float step_cache_ratio = (total_steps > 0) ? (static_cast<float>(cached_steps.size()) / total_steps * 100.0f) : 0.0f;
+
+        LOG_INFO("CacheDIT: steps_cached=%zu/%d (%.1f%%), blocks_cached=%d/%d (%.1f%%), accum_diff=%.4f",
+                 cached_steps.size(), total_steps, step_cache_ratio,
+                 total_blocks_cached, total_blocks, cache_ratio,
+                 accumulated_residual_diff);
+    }
+
+    std::string get_summary() const {
+        char buf[256];
+        snprintf(buf, sizeof(buf),
+                 "CacheDIT[thresh=%.2f]: cached %zu/%d steps, %d/%d blocks",
+                 config.dbcache.residual_diff_threshold,
+                 cached_steps.size(), total_steps,
+                 total_blocks_cached, total_blocks_computed + total_blocks_cached);
+        return std::string(buf);
+    }
+};
+
+inline std::vector<int> parse_scm_mask(const std::string& mask_str) {
+    std::vector<int> mask;
+    if (mask_str.empty())
+        return mask;
+
+    size_t pos   = 0;
+    size_t start = 0;
+    while ((pos = mask_str.find(',', start)) != std::string::npos) {
+        std::string token = mask_str.substr(start, pos - start);
+        mask.push_back(std::stoi(token));
+        start = pos + 1;
+    }
+    if (start < mask_str.length()) {
+        mask.push_back(std::stoi(mask_str.substr(start)));
+    }
+
+    return mask;
+}
+
+inline std::vector<int> generate_scm_mask(
+    const std::vector<int>& compute_bins,
+    const std::vector<int>& cache_bins,
+    int total_steps) {
+    std::vector<int> mask;
+    size_t c_idx = 0, cache_idx = 0;
+
+    while (static_cast<int>(mask.size()) < total_steps) {
+        if (c_idx < compute_bins.size()) {
+            for (int i = 0; i < compute_bins[c_idx] && static_cast<int>(mask.size()) < total_steps; i++) {
+                mask.push_back(1);
+            }
+            c_idx++;
+        }
+        if (cache_idx < cache_bins.size()) {
+            for (int i = 0; i < cache_bins[cache_idx] && static_cast<int>(mask.size()) < total_steps; i++) {
+                mask.push_back(0);
+            }
+            cache_idx++;
+        }
+        if (c_idx >= compute_bins.size() && cache_idx >= cache_bins.size())
+            break;
+    }
+
+    if (!mask.empty()) {
+        mask.back() = 1;
+    }
+
+    return mask;
+}
+
+inline std::vector<int> get_scm_preset(const std::string& preset, int total_steps) {
+    struct Preset {
+        std::vector<int> compute_bins;
+        std::vector<int> cache_bins;
+    };
+
+    Preset slow   = {{8, 3, 3, 2, 1, 1}, {1, 2, 2, 2, 3}};
+    Preset medium = {{6, 2, 2, 2, 2, 1}, {1, 3, 3, 3, 3}};
+    Preset fast   = {{6, 1, 1, 1, 1, 1}, {1, 3, 4, 5, 4}};
+    Preset ultra  = {{4, 1, 1, 1, 1}, {2, 5, 6, 7}};
+
+    Preset* p = nullptr;
+    if (preset == "slow" || preset == "s" || preset == "S")
+        p = &slow;
+    else if (preset == "medium" || preset == "m" || preset == "M")
+        p = &medium;
+    else if (preset == "fast" || preset == "f" || preset == "F")
+        p = &fast;
+    else if (preset == "ultra" || preset == "u" || preset == "U")
+        p = &ultra;
+    else
+        return {};
+
+    if (total_steps != 28 && total_steps > 0) {
+        float scale = static_cast<float>(total_steps) / 28.0f;
+        std::vector<int> scaled_compute, scaled_cache;
+
+        for (int v : p->compute_bins) {
+            scaled_compute.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
+        }
+        for (int v : p->cache_bins) {
+            scaled_cache.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
+        }
+
+        return generate_scm_mask(scaled_compute, scaled_cache, total_steps);
+    }
+
+    return generate_scm_mask(p->compute_bins, p->cache_bins, total_steps);
+}
+
+inline float get_preset_threshold(const std::string& preset) {
+    if (preset == "slow" || preset == "s" || preset == "S")
+        return 0.20f;
+    if (preset == "medium" || preset == "m" || preset == "M")
+        return 0.25f;
+    if (preset == "fast" || preset == "f" || preset == "F")
+        return 0.30f;
+    if (preset == "ultra" || preset == "u" || preset == "U")
+        return 0.34f;
+    return 0.08f;
+}
+
+inline int get_preset_warmup(const std::string& preset) {
+    if (preset == "slow" || preset == "s" || preset == "S")
+        return 8;
+    if (preset == "medium" || preset == "m" || preset == "M")
+        return 6;
+    if (preset == "fast" || preset == "f" || preset == "F")
+        return 6;
+    if (preset == "ultra" || preset == "u" || preset == "U")
+        return 4;
+    return 8;
+}
+
+inline int get_preset_Fn(const std::string& preset) {
+    if (preset == "slow" || preset == "s" || preset == "S")
+        return 8;
+    if (preset == "medium" || preset == "m" || preset == "M")
+        return 8;
+    if (preset == "fast" || preset == "f" || preset == "F")
+        return 6;
+    if (preset == "ultra" || preset == "u" || preset == "U")
+        return 4;
+    return 8;
+}
+
+inline int get_preset_Bn(const std::string& preset) {
+    (void)preset;
+    return 0;
+}
+
+inline void parse_dbcache_options(const std::string& opts, DBCacheConfig& cfg) {
+    if (opts.empty())
+        return;
+
+    int Fn = 8, Bn = 0, warmup = 8, max_cached = -1, max_cont = -1;
+    float thresh = 0.08f;
+
+    sscanf(opts.c_str(), "%d,%d,%f,%d,%d,%d",
+           &Fn, &Bn, &thresh, &warmup, &max_cached, &max_cont);
+
+    cfg.Fn_compute_blocks           = Fn;
+    cfg.Bn_compute_blocks           = Bn;
+    cfg.residual_diff_threshold     = thresh;
+    cfg.max_warmup_steps            = warmup;
+    cfg.max_cached_steps            = max_cached;
+    cfg.max_continuous_cached_steps = max_cont;
+}
+
+inline void parse_taylorseer_options(const std::string& opts, TaylorSeerConfig& cfg) {
+    if (opts.empty())
+        return;
+
+    int n_deriv = 1, warmup = 2, interval = 1;
+    sscanf(opts.c_str(), "%d,%d,%d", &n_deriv, &warmup, &interval);
+
+    cfg.n_derivatives       = n_deriv;
+    cfg.max_warmup_steps    = warmup;
+    cfg.skip_interval_steps = interval;
+}
+
+struct CacheDitConditionState {
+    DBCacheConfig config;
+    TaylorSeerConfig taylor_config;
+    bool initialized = false;
+
+    int current_step_index = -1;
+    bool step_active       = false;
+    bool skip_current_step = false;
+    bool initial_step      = true;
+    int warmup_remaining   = 0;
+    std::vector<int> cached_steps;
+    int continuous_cached_steps     = 0;
+    float accumulated_residual_diff = 0.0f;
+    int total_steps_skipped         = 0;
+
+    const void* anchor_condition = nullptr;
+
+    struct CacheEntry {
+        std::vector<float> diff;
+        std::vector<float> prev_input;
+        std::vector<float> prev_output;
+        bool has_prev = false;
+    };
+    std::unordered_map<const void*, CacheEntry> cache_diffs;
+
+    TaylorSeerState taylor_state;
+
+    float start_sigma = std::numeric_limits<float>::max();
+    float end_sigma   = 0.0f;
+
+    void reset_runtime() {
+        current_step_index = -1;
+        step_active        = false;
+        skip_current_step  = false;
+        initial_step       = true;
+        warmup_remaining   = config.max_warmup_steps;
+        cached_steps.clear();
+        continuous_cached_steps   = 0;
+        accumulated_residual_diff = 0.0f;
+        total_steps_skipped       = 0;
+        anchor_condition          = nullptr;
+        cache_diffs.clear();
+        taylor_state.reset();
+    }
+
+    void init(const DBCacheConfig& dbcfg, const TaylorSeerConfig& tcfg) {
+        config        = dbcfg;
+        taylor_config = tcfg;
+        initialized   = dbcfg.enabled || tcfg.enabled;
+        reset_runtime();
+
+        if (taylor_config.enabled) {
+            taylor_state.init(taylor_config.n_derivatives, 0);
+        }
+    }
+
+    void set_sigmas(const std::vector<float>& sigmas) {
+        if (!initialized || sigmas.size() < 2)
+            return;
+
+        float start_percent = 0.15f;
+        float end_percent   = 0.95f;
+
+        size_t n_steps    = sigmas.size() - 1;
+        size_t start_step = static_cast<size_t>(start_percent * n_steps);
+        size_t end_step   = static_cast<size_t>(end_percent * n_steps);
+
+        if (start_step >= n_steps)
+            start_step = n_steps - 1;
+        if (end_step >= n_steps)
+            end_step = n_steps - 1;
+
+        start_sigma = sigmas[start_step];
+        end_sigma   = sigmas[end_step];
+
+        if (start_sigma < end_sigma) {
+            std::swap(start_sigma, end_sigma);
+        }
+    }
+
+    bool enabled() const {
+        return initialized && (config.enabled || taylor_config.enabled);
+    }
+
+    void begin_step(int step_index, float sigma) {
+        if (!enabled())
+            return;
+        if (step_index == current_step_index)
+            return;
+
+        current_step_index = step_index;
+        skip_current_step  = false;
+        step_active        = false;
+
+        if (sigma > start_sigma)
+            return;
+        if (!(sigma > end_sigma))
+            return;
+
+        step_active = true;
+
+        if (warmup_remaining > 0) {
+            warmup_remaining--;
+            return;
+        }
+
+        if (!config.steps_computation_mask.empty()) {
+            if (step_index < static_cast<int>(config.steps_computation_mask.size())) {
+                if (config.steps_computation_mask[step_index] == 1) {
+                    return;
+                }
+            }
+        }
+
+        if (config.max_cached_steps >= 0 &&
+            static_cast<int>(cached_steps.size()) >= config.max_cached_steps) {
+            return;
+        }
+
+        if (config.max_continuous_cached_steps >= 0 &&
+            continuous_cached_steps >= config.max_continuous_cached_steps) {
+            return;
+        }
+    }
+
+    bool step_is_active() const {
+        return enabled() && step_active;
+    }
+
+    bool is_step_skipped() const {
+        return enabled() && step_active && skip_current_step;
+    }
+
+    bool has_cache(const void* cond) const {
+        auto it = cache_diffs.find(cond);
+        return it != cache_diffs.end() && !it->second.diff.empty();
+    }
+
+    void update_cache(const void* cond, const float* input, const float* output, size_t size) {
+        CacheEntry& entry = cache_diffs[cond];
+        entry.diff.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.diff[i] = output[i] - input[i];
+        }
+
+        entry.prev_input.resize(size);
+        entry.prev_output.resize(size);
+        for (size_t i = 0; i < size; i++) {
+            entry.prev_input[i]  = input[i];
+            entry.prev_output[i] = output[i];
+        }
+        entry.has_prev = true;
+    }
+
+    void apply_cache(const void* cond, const float* input, float* output, size_t size) {
+        auto it = cache_diffs.find(cond);
+        if (it == cache_diffs.end() || it->second.diff.empty())
+            return;
+        if (it->second.diff.size() != size)
+            return;
+
+        for (size_t i = 0; i < size; i++) {
+            output[i] = input[i] + it->second.diff[i];
+        }
+    }
+
+    bool before_condition(const void* cond, struct ggml_tensor* input, struct ggml_tensor* output, float sigma, int step_index) {
+        if (!enabled() || step_index < 0)
+            return false;
+
+        if (step_index != current_step_index) {
+            begin_step(step_index, sigma);
+        }
+
+        if (!step_active)
+            return false;
+
+        if (initial_step) {
+            anchor_condition = cond;
+            initial_step     = false;
+        }
+
+        bool is_anchor = (cond == anchor_condition);
+
+        if (skip_current_step) {
+            if (has_cache(cond)) {
+                apply_cache(cond, (float*)input->data, (float*)output->data,
+                            static_cast<size_t>(ggml_nelements(output)));
+                return true;
+            }
+            return false;
+        }
+
+        if (!is_anchor)
+            return false;
+
+        auto it = cache_diffs.find(cond);
+        if (it == cache_diffs.end() || !it->second.has_prev)
+            return false;
+
+        size_t ne = static_cast<size_t>(ggml_nelements(input));
+        if (it->second.prev_input.size() != ne)
+            return false;
+
+        float* input_data = (float*)input->data;
+        float diff        = CacheDitState::calculate_residual_diff(
+                   it->second.prev_input.data(), input_data, ne);
+
+        float effective_threshold = config.residual_diff_threshold;
+        if (config.Fn_compute_blocks > 0) {
+            float fn_confidence = 1.0f + 0.02f * (config.Fn_compute_blocks - 8);
+            fn_confidence       = std::max(0.5f, std::min(2.0f, fn_confidence));
+            effective_threshold *= fn_confidence;
+        }
+        if (config.Bn_compute_blocks > 0) {
+            float bn_quality = 1.0f - 0.03f * config.Bn_compute_blocks;
+            bn_quality       = std::max(0.5f, std::min(1.0f, bn_quality));
+            effective_threshold *= bn_quality;
+        }
+
+        if (diff < effective_threshold) {
+            skip_current_step = true;
+            total_steps_skipped++;
+            cached_steps.push_back(current_step_index);
+            continuous_cached_steps++;
+            accumulated_residual_diff += diff;
+            apply_cache(cond, input_data, (float*)output->data, ne);
+            return true;
+        }
+
+        continuous_cached_steps = 0;
+        return false;
+    }
+
+    void after_condition(const void* cond, struct ggml_tensor* input, struct ggml_tensor* output) {
+        if (!step_is_active())
+            return;
+
+        size_t ne = static_cast<size_t>(ggml_nelements(output));
+        update_cache(cond, (float*)input->data, (float*)output->data, ne);
+
+        if (cond == anchor_condition && taylor_config.enabled) {
+            taylor_state.update_derivatives((float*)output->data, ne, current_step_index);
+        }
+    }
+
+    void log_metrics() const {
+        if (!enabled())
+            return;
+
+        LOG_INFO("CacheDIT: steps_skipped=%d/%d (%.1f%%), accum_residual_diff=%.4f",
+                 total_steps_skipped,
+                 current_step_index + 1,
+                 (current_step_index > 0) ? (100.0f * total_steps_skipped / (current_step_index + 1)) : 0.0f,
+                 accumulated_residual_diff);
+    }
+};
+
+#endif

denoiser.hpp

@@ -869,7 +869,7 @@ static bool sample_k_diffusion(sample_method_t method,
 
             for (int i = 0; i < steps; i++) {
                 // denoise
-                ggml_tensor* denoised = model(x, sigmas[i], i + 1);
+                ggml_tensor* denoised = model(x, sigmas[i], -(i + 1));
                 if (denoised == nullptr) {
                     return false;
                 }
@@ -927,7 +927,7 @@ static bool sample_k_diffusion(sample_method_t method,
 
             for (int i = 0; i < steps; i++) {
                 // denoise
-                ggml_tensor* denoised = model(x, sigmas[i], i + 1);
+                ggml_tensor* denoised = model(x, sigmas[i], -(i + 1));
                 if (denoised == nullptr) {
                     return false;
                 }

diffusion_model.hpp

@@ -320,8 +320,9 @@ struct QwenImageModel : public DiffusionModel {
                    bool offload_params_to_cpu,
                    const String2TensorStorage& tensor_storage_map = {},
                    const std::string prefix                       = "model.diffusion_model",
-                   SDVersion version                              = VERSION_QWEN_IMAGE)
-        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_storage_map, prefix, version) {
+                   SDVersion version                              = VERSION_QWEN_IMAGE,
+                   bool zero_cond_t                               = false)
+        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_storage_map, prefix, version, zero_cond_t) {
     }
 
     std::string get_desc() override {

docs/caching.md

@@ -0,0 +1,126 @@
+## Caching
+
+Caching methods accelerate diffusion inference by reusing intermediate computations when changes between steps are small.
+
+### Cache Modes
+
+| Mode | Target | Description |
+|------|--------|-------------|
+| `ucache` | UNET models | Condition-level caching with error tracking |
+| `easycache` | DiT models | Condition-level cache |
+| `dbcache` | DiT models | Block-level L1 residual threshold |
+| `taylorseer` | DiT models | Taylor series approximation |
+| `cache-dit` | DiT models | Combined DBCache + TaylorSeer |
+
+### UCache (UNET Models)
+
+UCache caches the residual difference (output - input) and reuses it when input changes are below threshold.
+
+```bash
+sd-cli -m model.safetensors -p "a cat" --cache-mode ucache --cache-option "threshold=1.5"
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `threshold` | Error threshold for reuse decision | 1.0 |
+| `start` | Start caching at this percent of steps | 0.15 |
+| `end` | Stop caching at this percent of steps | 0.95 |
+| `decay` | Error decay rate (0-1) | 1.0 |
+| `relative` | Scale threshold by output norm (0/1) | 1 |
+| `reset` | Reset error after computing (0/1) | 1 |
+
+#### Reset Parameter
+
+The `reset` parameter controls error accumulation behavior:
+
+- `reset=1` (default): Resets accumulated error after each computed step. More aggressive caching, works well with most samplers.
+- `reset=0`: Keeps error accumulated. More conservative, recommended for `euler_a` sampler.
+
+### EasyCache (DiT Models)
+
+Condition-level caching for DiT models. Caches and reuses outputs when input changes are below threshold.
+
+```bash
+--cache-mode easycache --cache-option "threshold=0.3"
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `threshold` | Input change threshold for reuse | 0.2 |
+| `start` | Start caching at this percent of steps | 0.15 |
+| `end` | Stop caching at this percent of steps | 0.95 |
+
+### Cache-DIT (DiT Models)
+
+For DiT models like FLUX and QWEN, use block-level caching modes.
+
+#### DBCache
+
+Caches blocks based on L1 residual difference threshold:
+
+```bash
+--cache-mode dbcache --cache-option "threshold=0.25,warmup=4"
+```
+
+#### TaylorSeer
+
+Uses Taylor series approximation to predict block outputs:
+
+```bash
+--cache-mode taylorseer
+```
+
+#### Cache-DIT (Combined)
+
+Combines DBCache and TaylorSeer:
+
+```bash
+--cache-mode cache-dit --cache-preset fast
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `Fn` | Front blocks to always compute | 8 |
+| `Bn` | Back blocks to always compute | 0 |
+| `threshold` | L1 residual difference threshold | 0.08 |
+| `warmup` | Steps before caching starts | 8 |
+
+#### Presets
+
+Available presets: `slow`, `medium`, `fast`, `ultra` (or `s`, `m`, `f`, `u`).
+
+```bash
+--cache-mode cache-dit --cache-preset fast
+```
+
+#### SCM Options
+
+Steps Computation Mask controls which steps can be cached:
+
+```bash
+--scm-mask "1,1,1,1,0,0,1,0,0,0,1,0,0,0,1,0,0,0,1,1"
+```
+
+Mask values: `1` = compute, `0` = can cache.
+
+| Policy | Description |
+|--------|-------------|
+| `dynamic` | Check threshold before caching |
+| `static` | Always cache on cacheable steps |
+
+```bash
+--scm-policy dynamic
+```
+
+### Performance Tips
+
+- Start with default thresholds and adjust based on output quality
+- Lower threshold = better quality, less speedup
+- Higher threshold = more speedup, potential quality loss
+- More steps generally means more caching opportunities

docs/qwen_image_edit.md

@@ -9,6 +9,9 @@
     - Qwen Image Edit 2509
         - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
         - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/tree/main
+    - Qwen Image Edit 2511
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/tree/main
 - Download vae
     - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
 - Download qwen_2.5_vl 7b
@@ -32,4 +35,14 @@
 .\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --llm_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
 ```
 
-<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />
+<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />
+
+### Qwen Image Edit 2511
+
+To use the new Qwen Image Edit 2511 mode, the  `--qwen-image-zero-cond-t` flag must be enabled; otherwise, image editing quality will degrade significantly.
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-edit-2511-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'"  --qwen-image-zero-cond-t
+```
+
+<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2511.png" />

examples/cli/README.md

@@ -4,7 +4,8 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>       path to write result image to (default: ./output.png)
+  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png)
+  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
   --preview-path <string>     path to write preview image to (default: ./preview.png)
   --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
                               every step)
@@ -47,12 +48,16 @@ Context Options:
   --vae-tiling                             process vae in tiles to reduce memory usage
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
   --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  --mmap                                   whether to memory-map model
   --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
   --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
+  --circularx                              enable circular RoPE wrapping on x-axis (width) only
+  --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --chroma-disable-dit-mask                disable dit mask for chroma
   --chroma-enable-t5-mask                  enable t5 mask for chroma
   --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
@@ -94,6 +99,7 @@ Generation Options:
   --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
                                            NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
+  --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
   --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
@@ -121,11 +127,17 @@ Generation Options:
                                            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
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm],
-                                           default: discrete
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
+                                           kl_optimal, lcm], 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])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --easycache                              enable EasyCache for DiT models with optional "threshold,start_percent,end_percent" (default: 0.2,0.15,0.95)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)
+  --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
+                                           "threshold=0.25" or "threshold=1.5,reset=0"
+  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+  --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
 ```

examples/cli/main.cpp

@@ -26,9 +26,12 @@ const char* previews_str[] = {
     "vae",
 };
 
+std::regex format_specifier_regex("(?:[^%]|^)(?:%%)*(%\\d{0,3}d)");
+
 struct SDCliParams {
     SDMode mode             = IMG_GEN;
     std::string output_path = "output.png";
+    int output_begin_idx    = -1;
 
     bool verbose          = false;
     bool canny_preprocess = false;
@@ -50,7 +53,7 @@ struct SDCliParams {
         options.string_options = {
             {"-o",
              "--output",
-             "path to write result image to (default: ./output.png)",
+             "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png)",
              &output_path},
             {"",
              "--preview-path",
@@ -63,6 +66,10 @@ struct SDCliParams {
              "--preview-interval",
              "interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at every step)",
              &preview_interval},
+            {"",
+             "--output-begin-idx",
+             "starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)",
+             &output_begin_idx},
         };
 
         options.bool_options = {
@@ -73,7 +80,7 @@ struct SDCliParams {
             {"",
              "--convert-name",
              "convert tensor name (for convert mode)",
-             true, &canny_preprocess},
+             true, &convert_name},
             {"-v",
              "--verbose",
              "print extra info",
@@ -344,6 +351,114 @@ void step_callback(int step, int frame_count, sd_image_t* image, bool is_noisy,
     }
 }
 
+std::string format_frame_idx(std::string pattern, int frame_idx) {
+    std::smatch match;
+    std::string result = pattern;
+    while (std::regex_search(result, match, format_specifier_regex)) {
+        std::string specifier = match.str(1);
+        char buffer[32];
+        snprintf(buffer, sizeof(buffer), specifier.c_str(), frame_idx);
+        result.replace(match.position(1), match.length(1), buffer);
+    }
+
+    // Then replace all '%%' with '%'
+    size_t pos = 0;
+    while ((pos = result.find("%%", pos)) != std::string::npos) {
+        result.replace(pos, 2, "%");
+        pos += 1;
+    }
+    return result;
+}
+
+bool save_results(const SDCliParams& cli_params,
+                  const SDContextParams& ctx_params,
+                  const SDGenerationParams& gen_params,
+                  sd_image_t* results,
+                  int num_results) {
+    if (results == nullptr || num_results <= 0) {
+        return false;
+    }
+
+    namespace fs      = std::filesystem;
+    fs::path out_path = cli_params.output_path;
+
+    if (!out_path.parent_path().empty()) {
+        std::error_code ec;
+        fs::create_directories(out_path.parent_path(), ec);
+        if (ec) {
+            LOG_ERROR("failed to create directory '%s': %s",
+                      out_path.parent_path().string().c_str(), ec.message().c_str());
+            return false;
+        }
+    }
+
+    fs::path base_path = out_path;
+    fs::path ext       = out_path.has_extension() ? out_path.extension() : fs::path{};
+    if (!ext.empty())
+        base_path.replace_extension();
+
+    std::string ext_lower = ext.string();
+    std::transform(ext_lower.begin(), ext_lower.end(), ext_lower.begin(), ::tolower);
+    bool is_jpg = (ext_lower == ".jpg" || ext_lower == ".jpeg" || ext_lower == ".jpe");
+
+    int output_begin_idx = cli_params.output_begin_idx;
+    if (output_begin_idx < 0) {
+        output_begin_idx = 0;
+    }
+
+    auto write_image = [&](const fs::path& path, int idx) {
+        const sd_image_t& img = results[idx];
+        if (!img.data)
+            return;
+
+        std::string params = get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + idx);
+        int ok             = 0;
+        if (is_jpg) {
+            ok = stbi_write_jpg(path.string().c_str(), img.width, img.height, img.channel, img.data, 90, params.c_str());
+        } else {
+            ok = stbi_write_png(path.string().c_str(), img.width, img.height, img.channel, img.data, 0, params.c_str());
+        }
+        LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
+    };
+
+    if (std::regex_search(cli_params.output_path, format_specifier_regex)) {
+        if (!is_jpg && ext_lower != ".png")
+            ext = ".png";
+        fs::path pattern = base_path;
+        pattern += ext;
+
+        for (int i = 0; i < num_results; ++i) {
+            fs::path img_path = format_frame_idx(pattern.string(), output_begin_idx + i);
+            write_image(img_path, i);
+        }
+        return true;
+    }
+
+    if (cli_params.mode == VID_GEN && num_results > 1) {
+        if (ext_lower != ".avi")
+            ext = ".avi";
+        fs::path video_path = base_path;
+        video_path += ext;
+        create_mjpg_avi_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps);
+        LOG_INFO("save result MJPG AVI video to '%s'", video_path.string().c_str());
+        return true;
+    }
+
+    if (!is_jpg && ext_lower != ".png")
+        ext = ".png";
+
+    for (int i = 0; i < num_results; ++i) {
+        fs::path img_path = base_path;
+        if (num_results > 1) {
+            img_path += "_" + std::to_string(output_begin_idx + i);
+        }
+        img_path += ext;
+        write_image(img_path, i);
+    }
+
+    return true;
+}
+
 int main(int argc, const char* argv[]) {
     if (argc > 1 && std::string(argv[1]) == "--version") {
         std::cout << version_string() << "\n";
@@ -617,7 +732,7 @@ int main(int argc, const char* argv[]) {
                     gen_params.pm_style_strength,
                 },  // pm_params
                 ctx_params.vae_tiling_params,
-                gen_params.easycache_params,
+                gen_params.cache_params,
             };
 
             results     = generate_image(sd_ctx, &img_gen_params);
@@ -642,7 +757,7 @@ int main(int argc, const char* argv[]) {
                 gen_params.seed,
                 gen_params.video_frames,
                 gen_params.vace_strength,
-                gen_params.easycache_params,
+                gen_params.cache_params,
             };
 
             results = generate_video(sd_ctx, &vid_gen_params, &num_results);
@@ -687,67 +802,8 @@ int main(int argc, const char* argv[]) {
         }
     }
 
-    // create directory if not exists
-    {
-        const fs::path out_path = cli_params.output_path;
-        if (const fs::path out_dir = out_path.parent_path(); !out_dir.empty()) {
-            std::error_code ec;
-            fs::create_directories(out_dir, ec);  // OK if already exists
-            if (ec) {
-                LOG_ERROR("failed to create directory '%s': %s",
-                          out_dir.string().c_str(), ec.message().c_str());
-                return 1;
-            }
-        }
-    }
-
-    std::string base_path;
-    std::string file_ext;
-    std::string file_ext_lower;
-    bool is_jpg;
-    size_t last_dot_pos   = cli_params.output_path.find_last_of(".");
-    size_t last_slash_pos = std::min(cli_params.output_path.find_last_of("/"),
-                                     cli_params.output_path.find_last_of("\\"));
-    if (last_dot_pos != std::string::npos && (last_slash_pos == std::string::npos || last_dot_pos > last_slash_pos)) {  // filename has extension
-        base_path = cli_params.output_path.substr(0, last_dot_pos);
-        file_ext = file_ext_lower = cli_params.output_path.substr(last_dot_pos);
-        std::transform(file_ext.begin(), file_ext.end(), file_ext_lower.begin(), ::tolower);
-        is_jpg = (file_ext_lower == ".jpg" || file_ext_lower == ".jpeg" || file_ext_lower == ".jpe");
-    } else {
-        base_path = cli_params.output_path;
-        file_ext = file_ext_lower = "";
-        is_jpg                    = false;
-    }
-
-    if (cli_params.mode == VID_GEN && num_results > 1) {
-        std::string vid_output_path = cli_params.output_path;
-        if (file_ext_lower == ".png") {
-            vid_output_path = base_path + ".avi";
-        }
-        create_mjpg_avi_from_sd_images(vid_output_path.c_str(), results, num_results, gen_params.fps);
-        LOG_INFO("save result MJPG AVI video to '%s'\n", vid_output_path.c_str());
-    } else {
-        // appending ".png" to absent or unknown extension
-        if (!is_jpg && file_ext_lower != ".png") {
-            base_path += file_ext;
-            file_ext = ".png";
-        }
-        for (int i = 0; i < num_results; i++) {
-            if (results[i].data == nullptr) {
-                continue;
-            }
-            int write_ok;
-            std::string final_image_path = i > 0 ? base_path + "_" + std::to_string(i + 1) + file_ext : base_path + file_ext;
-            if (is_jpg) {
-                write_ok = stbi_write_jpg(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
-                                          results[i].data, 90, get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + i).c_str());
-                LOG_INFO("save result JPEG image to '%s' (%s)", final_image_path.c_str(), write_ok == 0 ? "failure" : "success");
-            } else {
-                write_ok = stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
-                                          results[i].data, 0, get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + i).c_str());
-                LOG_INFO("save result PNG image to '%s' (%s)", final_image_path.c_str(), write_ok == 0 ? "failure" : "success");
-            }
-        }
+    if (!save_results(cli_params, ctx_params, gen_params, results, num_results)) {
+        return 1;
     }
 
     for (int i = 0; i < num_results; i++) {

examples/common/common.hpp

@@ -95,17 +95,28 @@ static void print_utf8(FILE* stream, const char* utf8) {
                    ? GetStdHandle(STD_ERROR_HANDLE)
                    : GetStdHandle(STD_OUTPUT_HANDLE);
 
-    int wlen = MultiByteToWideChar(CP_UTF8, 0, utf8, -1, NULL, 0);
-    if (wlen <= 0)
-        return;
+    DWORD mode;
+    BOOL is_console = GetConsoleMode(h, &mode);
 
-    wchar_t* wbuf = (wchar_t*)malloc(wlen * sizeof(wchar_t));
-    MultiByteToWideChar(CP_UTF8, 0, utf8, -1, wbuf, wlen);
+    if (is_console) {
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, utf8, -1, NULL, 0);
+        if (wlen <= 0)
+            return;
 
-    DWORD written;
-    WriteConsoleW(h, wbuf, wlen - 1, &written, NULL);
+        wchar_t* wbuf = (wchar_t*)malloc(wlen * sizeof(wchar_t));
+        if (!wbuf)
+            return;
 
-    free(wbuf);
+        MultiByteToWideChar(CP_UTF8, 0, utf8, -1, wbuf, wlen);
+
+        DWORD written;
+        WriteConsoleW(h, wbuf, wlen - 1, &written, NULL);
+
+        free(wbuf);
+    } else {
+        DWORD written;
+        WriteFile(h, utf8, (DWORD)strlen(utf8), &written, NULL);
+    }
 #else
     fputs(utf8, stream);
 #endif
@@ -442,6 +453,7 @@ struct SDContextParams {
     rng_type_t rng_type         = CUDA_RNG;
     rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
     bool offload_params_to_cpu  = false;
+    bool enable_mmap            = false;
     bool control_net_cpu        = false;
     bool clip_on_cpu            = false;
     bool vae_on_cpu             = false;
@@ -457,6 +469,8 @@ struct SDContextParams {
     bool chroma_use_t5_mask  = false;
     int chroma_t5_mask_pad   = 1;
 
+    bool qwen_image_zero_cond_t = false;
+
     prediction_t prediction           = PREDICTION_COUNT;
     lora_apply_mode_t lora_apply_mode = LORA_APPLY_AUTO;
 
@@ -585,6 +599,10 @@ struct SDContextParams {
              "--offload-to-cpu",
              "place the weights in RAM to save VRAM, and automatically load them into VRAM when needed",
              true, &offload_params_to_cpu},
+            {"",
+             "--mmap",
+             "whether to memory-map model",
+             true, &enable_mmap},
             {"",
              "--control-net-cpu",
              "keep controlnet in cpu (for low vram)",
@@ -625,6 +643,10 @@ struct SDContextParams {
              "--chroma-disable-dit-mask",
              "disable dit mask for chroma",
              false, &chroma_use_dit_mask},
+            {"",
+             "--qwen-image-zero-cond-t",
+             "enable zero_cond_t for qwen image",
+             true, &qwen_image_zero_cond_t},
             {"",
              "--chroma-enable-t5-mask",
              "enable t5 mask for chroma",
@@ -878,6 +900,7 @@ struct SDContextParams {
             << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
             << "  flow_shift: " << (std::isinf(flow_shift) ? "INF" : std::to_string(flow_shift)) << "\n"
             << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
+            << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
             << "  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"
@@ -888,6 +911,7 @@ struct SDContextParams {
             << "  circular_x: " << (circular_x ? "true" : "false") << ",\n"
             << "  circular_y: " << (circular_y ? "true" : "false") << ",\n"
             << "  chroma_use_dit_mask: " << (chroma_use_dit_mask ? "true" : "false") << ",\n"
+            << "  qwen_image_zero_cond_t: " << (qwen_image_zero_cond_t ? "true" : "false") << ",\n"
             << "  chroma_use_t5_mask: " << (chroma_use_t5_mask ? "true" : "false") << ",\n"
             << "  chroma_t5_mask_pad: " << chroma_t5_mask_pad << ",\n"
             << "  prediction: " << sd_prediction_name(prediction) << ",\n"
@@ -940,6 +964,7 @@ struct SDContextParams {
             prediction,
             lora_apply_mode,
             offload_params_to_cpu,
+            enable_mmap,
             clip_on_cpu,
             control_net_cpu,
             vae_on_cpu,
@@ -953,6 +978,7 @@ struct SDContextParams {
             chroma_use_dit_mask,
             chroma_use_t5_mask,
             chroma_t5_mask_pad,
+            qwen_image_zero_cond_t,
             flow_shift,
         };
         return sd_ctx_params;
@@ -1018,8 +1044,12 @@ struct SDGenerationParams {
 
     std::vector<float> custom_sigmas;
 
-    std::string easycache_option;
-    sd_easycache_params_t easycache_params;
+    std::string cache_mode;
+    std::string cache_option;
+    std::string cache_preset;
+    std::string scm_mask;
+    bool scm_policy_dynamic = true;
+    sd_cache_params_t cache_params{};
 
     float moe_boundary  = 0.875f;
     int video_frames    = 1;
@@ -1381,36 +1411,64 @@ struct SDGenerationParams {
             return 1;
         };
 
-        auto on_easycache_arg = [&](int argc, const char** argv, int index) {
-            const std::string default_values = "0.2,0.15,0.95";
-            auto looks_like_value            = [](const std::string& token) {
-                if (token.empty()) {
-                    return false;
-                }
-                if (token[0] != '-') {
-                    return true;
-                }
-                if (token.size() == 1) {
-                    return false;
-                }
-                unsigned char next = static_cast<unsigned char>(token[1]);
-                return std::isdigit(next) || token[1] == '.';
-            };
+        auto on_cache_mode_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
+            }
+            cache_mode = argv_to_utf8(index, argv);
+            if (cache_mode != "easycache" && cache_mode != "ucache" &&
+                cache_mode != "dbcache" && cache_mode != "taylorseer" && cache_mode != "cache-dit") {
+                fprintf(stderr, "error: invalid cache mode '%s', must be 'easycache', 'ucache', 'dbcache', 'taylorseer', or 'cache-dit'\n", cache_mode.c_str());
+                return -1;
+            }
+            return 1;
+        };
 
-            std::string option_value;
-            int consumed = 0;
-            if (index + 1 < argc) {
-                std::string next_arg = argv[index + 1];
-                if (looks_like_value(next_arg)) {
-                    option_value = argv_to_utf8(index + 1, argv);
-                    consumed     = 1;
-                }
+        auto on_cache_option_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
             }
-            if (option_value.empty()) {
-                option_value = default_values;
+            cache_option = argv_to_utf8(index, argv);
+            return 1;
+        };
+
+        auto on_scm_mask_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
             }
-            easycache_option = option_value;
-            return consumed;
+            scm_mask = argv_to_utf8(index, argv);
+            return 1;
+        };
+
+        auto on_scm_policy_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
+            }
+            std::string policy = argv_to_utf8(index, argv);
+            if (policy == "dynamic") {
+                scm_policy_dynamic = true;
+            } else if (policy == "static") {
+                scm_policy_dynamic = false;
+            } else {
+                fprintf(stderr, "error: invalid scm policy '%s', must be 'dynamic' or 'static'\n", policy.c_str());
+                return -1;
+            }
+            return 1;
+        };
+
+        auto on_cache_preset_arg = [&](int argc, const char** argv, int index) {
+            if (++index >= argc) {
+                return -1;
+            }
+            cache_preset = argv_to_utf8(index, argv);
+            if (cache_preset != "slow" && cache_preset != "s" && cache_preset != "S" &&
+                cache_preset != "medium" && cache_preset != "m" && cache_preset != "M" &&
+                cache_preset != "fast" && cache_preset != "f" && cache_preset != "F" &&
+                cache_preset != "ultra" && cache_preset != "u" && cache_preset != "U") {
+                fprintf(stderr, "error: invalid cache preset '%s', must be 'slow'/'s', 'medium'/'m', 'fast'/'f', or 'ultra'/'u'\n", cache_preset.c_str());
+                return -1;
+            }
+            return 1;
         };
 
         options.manual_options = {
@@ -1449,9 +1507,25 @@ struct SDGenerationParams {
              "reference image for Flux Kontext models (can be used multiple times)",
              on_ref_image_arg},
             {"",
-             "--easycache",
-             "enable EasyCache for DiT models with optional \"threshold,start_percent,end_percent\" (default: 0.2,0.15,0.95)",
-             on_easycache_arg},
+             "--cache-mode",
+             "caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)",
+             on_cache_mode_arg},
+            {"",
+             "--cache-option",
+             "named cache params (key=value format, comma-separated). easycache/ucache: threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples: \"threshold=0.25\" or \"threshold=1.5,reset=0\"",
+             on_cache_option_arg},
+            {"",
+             "--cache-preset",
+             "cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'",
+             on_cache_preset_arg},
+            {"",
+             "--scm-mask",
+             "SCM steps mask for cache-dit: comma-separated 0/1 (e.g., \"1,1,1,0,0,1,0,0,1,0\") - 1=compute, 0=can cache",
+             on_scm_mask_arg},
+            {"",
+             "--scm-policy",
+             "SCM policy: 'dynamic' (default) or 'static'",
+             on_scm_policy_arg},
 
         };
 
@@ -1494,7 +1568,10 @@ struct SDGenerationParams {
 
         load_if_exists("prompt", prompt);
         load_if_exists("negative_prompt", negative_prompt);
-        load_if_exists("easycache_option", easycache_option);
+        load_if_exists("cache_mode", cache_mode);
+        load_if_exists("cache_option", cache_option);
+        load_if_exists("cache_preset", cache_preset);
+        load_if_exists("scm_mask", scm_mask);
 
         load_if_exists("clip_skip", clip_skip);
         load_if_exists("width", width);
@@ -1634,57 +1711,118 @@ struct SDGenerationParams {
             return false;
         }
 
-        if (!easycache_option.empty()) {
-            float values[3] = {0.0f, 0.0f, 0.0f};
-            std::stringstream ss(easycache_option);
+        sd_cache_params_init(&cache_params);
+
+        auto parse_named_params = [&](const std::string& opt_str) -> bool {
+            std::stringstream ss(opt_str);
             std::string token;
-            int idx = 0;
             while (std::getline(ss, token, ',')) {
-                auto trim = [](std::string& s) {
-                    const char* whitespace = " \t\r\n";
-                    auto start             = s.find_first_not_of(whitespace);
-                    if (start == std::string::npos) {
-                        s.clear();
-                        return;
-                    }
-                    auto end = s.find_last_not_of(whitespace);
-                    s        = s.substr(start, end - start + 1);
-                };
-                trim(token);
-                if (token.empty()) {
-                    LOG_ERROR("error: invalid easycache option '%s'", easycache_option.c_str());
-                    return false;
-                }
-                if (idx >= 3) {
-                    LOG_ERROR("error: easycache expects exactly 3 comma-separated values (threshold,start,end)\n");
+                size_t eq_pos = token.find('=');
+                if (eq_pos == std::string::npos) {
+                    LOG_ERROR("error: cache option '%s' missing '=' separator", token.c_str());
                     return false;
                 }
+                std::string key = token.substr(0, eq_pos);
+                std::string val = token.substr(eq_pos + 1);
                 try {
-                    values[idx] = std::stof(token);
+                    if (key == "threshold") {
+                        if (cache_mode == "easycache" || cache_mode == "ucache") {
+                            cache_params.reuse_threshold = std::stof(val);
+                        } else {
+                            cache_params.residual_diff_threshold = std::stof(val);
+                        }
+                    } else if (key == "start") {
+                        cache_params.start_percent = std::stof(val);
+                    } else if (key == "end") {
+                        cache_params.end_percent = std::stof(val);
+                    } else if (key == "decay") {
+                        cache_params.error_decay_rate = std::stof(val);
+                    } else if (key == "relative") {
+                        cache_params.use_relative_threshold = (std::stof(val) != 0.0f);
+                    } else if (key == "reset") {
+                        cache_params.reset_error_on_compute = (std::stof(val) != 0.0f);
+                    } else if (key == "Fn" || key == "fn") {
+                        cache_params.Fn_compute_blocks = std::stoi(val);
+                    } else if (key == "Bn" || key == "bn") {
+                        cache_params.Bn_compute_blocks = std::stoi(val);
+                    } else if (key == "warmup") {
+                        cache_params.max_warmup_steps = std::stoi(val);
+                    } else {
+                        LOG_ERROR("error: unknown cache parameter '%s'", key.c_str());
+                        return false;
+                    }
                 } catch (const std::exception&) {
-                    LOG_ERROR("error: invalid easycache value '%s'", token.c_str());
+                    LOG_ERROR("error: invalid value '%s' for parameter '%s'", val.c_str(), key.c_str());
                     return false;
                 }
-                idx++;
             }
-            if (idx != 3) {
-                LOG_ERROR("error: easycache expects exactly 3 comma-separated values (threshold,start,end)\n");
-                return false;
+            return true;
+        };
+
+        if (!cache_mode.empty()) {
+            if (cache_mode == "easycache") {
+                cache_params.mode                   = SD_CACHE_EASYCACHE;
+                cache_params.reuse_threshold        = 0.2f;
+                cache_params.start_percent          = 0.15f;
+                cache_params.end_percent            = 0.95f;
+                cache_params.error_decay_rate       = 1.0f;
+                cache_params.use_relative_threshold = true;
+                cache_params.reset_error_on_compute = true;
+            } else if (cache_mode == "ucache") {
+                cache_params.mode                   = SD_CACHE_UCACHE;
+                cache_params.reuse_threshold        = 1.0f;
+                cache_params.start_percent          = 0.15f;
+                cache_params.end_percent            = 0.95f;
+                cache_params.error_decay_rate       = 1.0f;
+                cache_params.use_relative_threshold = true;
+                cache_params.reset_error_on_compute = true;
+            } else if (cache_mode == "dbcache") {
+                cache_params.mode                    = SD_CACHE_DBCACHE;
+                cache_params.Fn_compute_blocks       = 8;
+                cache_params.Bn_compute_blocks       = 0;
+                cache_params.residual_diff_threshold = 0.08f;
+                cache_params.max_warmup_steps        = 8;
+            } else if (cache_mode == "taylorseer") {
+                cache_params.mode                    = SD_CACHE_TAYLORSEER;
+                cache_params.Fn_compute_blocks       = 8;
+                cache_params.Bn_compute_blocks       = 0;
+                cache_params.residual_diff_threshold = 0.08f;
+                cache_params.max_warmup_steps        = 8;
+            } else if (cache_mode == "cache-dit") {
+                cache_params.mode                    = SD_CACHE_CACHE_DIT;
+                cache_params.Fn_compute_blocks       = 8;
+                cache_params.Bn_compute_blocks       = 0;
+                cache_params.residual_diff_threshold = 0.08f;
+                cache_params.max_warmup_steps        = 8;
             }
-            if (values[0] < 0.0f) {
-                LOG_ERROR("error: easycache threshold must be non-negative\n");
-                return false;
+
+            if (!cache_option.empty()) {
+                if (!parse_named_params(cache_option)) {
+                    return false;
+                }
             }
-            if (values[1] < 0.0f || values[1] >= 1.0f || values[2] <= 0.0f || values[2] > 1.0f || values[1] >= values[2]) {
-                LOG_ERROR("error: easycache start/end percents must satisfy 0.0 <= start < end <= 1.0\n");
-                return false;
+
+            if (cache_mode == "easycache" || cache_mode == "ucache") {
+                if (cache_params.reuse_threshold < 0.0f) {
+                    LOG_ERROR("error: cache threshold must be non-negative");
+                    return false;
+                }
+                if (cache_params.start_percent < 0.0f || cache_params.start_percent >= 1.0f ||
+                    cache_params.end_percent <= 0.0f || cache_params.end_percent > 1.0f ||
+                    cache_params.start_percent >= cache_params.end_percent) {
+                    LOG_ERROR("error: cache start/end percents must satisfy 0.0 <= start < end <= 1.0");
+                    return false;
+                }
             }
-            easycache_params.enabled         = true;
-            easycache_params.reuse_threshold = values[0];
-            easycache_params.start_percent   = values[1];
-            easycache_params.end_percent     = values[2];
-        } else {
-            easycache_params.enabled = false;
+        }
+
+        if (cache_params.mode == SD_CACHE_DBCACHE ||
+            cache_params.mode == SD_CACHE_TAYLORSEER ||
+            cache_params.mode == SD_CACHE_CACHE_DIT) {
+            if (!scm_mask.empty()) {
+                cache_params.scm_mask = scm_mask.c_str();
+            }
+            cache_params.scm_policy_dynamic = scm_policy_dynamic;
         }
 
         sample_params.guidance.slg.layers                 = skip_layers.data();
@@ -1786,12 +1924,13 @@ struct SDGenerationParams {
             << "  high_noise_skip_layers: " << vec_to_string(high_noise_skip_layers) << ",\n"
             << "  high_noise_sample_params: " << high_noise_sample_params_str << ",\n"
             << "  custom_sigmas: " << vec_to_string(custom_sigmas) << ",\n"
-            << "  easycache_option: \"" << easycache_option << "\",\n"
-            << "  easycache: "
-            << (easycache_params.enabled ? "enabled" : "disabled")
-            << " (threshold=" << easycache_params.reuse_threshold
-            << ", start=" << easycache_params.start_percent
-            << ", end=" << easycache_params.end_percent << "),\n"
+            << "  cache_mode: \"" << cache_mode << "\",\n"
+            << "  cache_option: \"" << cache_option << "\",\n"
+            << "  cache: "
+            << (cache_params.mode != SD_CACHE_DISABLED ? "enabled" : "disabled")
+            << " (threshold=" << cache_params.reuse_threshold
+            << ", start=" << cache_params.start_percent
+            << ", end=" << cache_params.end_percent << "),\n"
             << "  moe_boundary: " << moe_boundary << ",\n"
             << "  video_frames: " << video_frames << ",\n"
             << "  fps: " << fps << ",\n"

examples/server/README.md

@@ -6,6 +6,7 @@ usage: ./bin/sd-server  [options]
 Svr Options:
   -l, --listen-ip <string>    server listen ip (default: 127.0.0.1)
   --listen-port <int>         server listen port (default: 1234)
+  --serve-html-path <string>  path to HTML file to serve at root (optional)
   -v, --verbose               print extra info
   --color                     colors the logging tags according to level
   -h, --help                  show this help message and exit
@@ -42,9 +43,13 @@ 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)
+  --mmap                                   whether to memory-map model
   --diffusion-fa                           use flash attention in the diffusion model
   --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
+  --circularx                              enable circular RoPE wrapping on x-axis (width) only
+  --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --chroma-disable-dit-mask                disable dit mask for chroma
   --chroma-enable-t5-mask                  enable t5 mask for chroma
   --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
@@ -114,11 +119,17 @@ Default Generation Options:
                                            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
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm],
-                                           default: discrete
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
+                                           kl_optimal, lcm], 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])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --easycache                              enable EasyCache for DiT models with optional "threshold,start_percent,end_percent" (default: 0.2,0.15,0.95)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)
+  --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
+                                           "threshold=0.25" or "threshold=1.5,reset=0"
+  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+  --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
 ```

examples/server/main.cpp

@@ -104,9 +104,10 @@ std::string iso_timestamp_now() {
 struct SDSvrParams {
     std::string listen_ip = "127.0.0.1";
     int listen_port       = 1234;
-    bool normal_exit      = false;
-    bool verbose          = false;
-    bool color            = false;
+    std::string serve_html_path;
+    bool normal_exit = false;
+    bool verbose     = false;
+    bool color       = false;
 
     ArgOptions get_options() {
         ArgOptions options;
@@ -115,7 +116,11 @@ struct SDSvrParams {
             {"-l",
              "--listen-ip",
              "server listen ip (default: 127.0.0.1)",
-             &listen_ip}};
+             &listen_ip},
+            {"",
+             "--serve-html-path",
+             "path to HTML file to serve at root (optional)",
+             &serve_html_path}};
 
         options.int_options = {
             {"",
@@ -159,6 +164,11 @@ struct SDSvrParams {
             LOG_ERROR("error: listen_port should be in the range [0, 65535]");
             return false;
         }
+
+        if (!serve_html_path.empty() && !fs::exists(serve_html_path)) {
+            LOG_ERROR("error: serve_html_path file does not exist: %s", serve_html_path.c_str());
+            return false;
+        }
         return true;
     }
 
@@ -167,6 +177,7 @@ struct SDSvrParams {
         oss << "SDSvrParams {\n"
             << "  listen_ip: " << listen_ip << ",\n"
             << "  listen_port: \"" << listen_port << "\",\n"
+            << "  serve_html_path: \"" << serve_html_path << "\",\n"
             << "}";
         return oss.str();
     }
@@ -191,12 +202,18 @@ void parse_args(int argc, const char** argv, SDSvrParams& svr_params, SDContextP
         exit(svr_params.normal_exit ? 0 : 1);
     }
 
+    const bool random_seed_requested = default_gen_params.seed < 0;
+
     if (!svr_params.process_and_check() ||
         !ctx_params.process_and_check(IMG_GEN) ||
         !default_gen_params.process_and_check(IMG_GEN, ctx_params.lora_model_dir)) {
         print_usage(argc, argv, options_vec);
         exit(1);
     }
+
+    if (random_seed_requested) {
+        default_gen_params.seed = -1;
+    }
 }
 
 std::string extract_and_remove_sd_cpp_extra_args(std::string& text) {
@@ -312,7 +329,18 @@ int main(int argc, const char** argv) {
 
     // health
     svr.Get("/", [&](const httplib::Request&, httplib::Response& res) {
-        res.set_content(R"({"ok":true,"service":"sd-cpp-http"})", "application/json");
+        if (!svr_params.serve_html_path.empty()) {
+            std::ifstream file(svr_params.serve_html_path);
+            if (file) {
+                std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+                res.set_content(content, "text/html");
+            } else {
+                res.status = 500;
+                res.set_content("Error: Unable to read HTML file", "text/plain");
+            }
+        } else {
+            res.set_content("Stable Diffusion Server is running", "text/plain");
+        }
     });
 
     // models endpoint (minimal)
@@ -432,7 +460,7 @@ int main(int argc, const char** argv) {
                     gen_params.pm_style_strength,
                 },  // pm_params
                 ctx_params.vae_tiling_params,
-                gen_params.easycache_params,
+                gen_params.cache_params,
             };
 
             sd_image_t* results = nullptr;
@@ -645,7 +673,7 @@ int main(int argc, const char** argv) {
                     gen_params.pm_style_strength,
                 },  // pm_params
                 ctx_params.vae_tiling_params,
-                gen_params.easycache_params,
+                gen_params.cache_params,
             };
 
             sd_image_t* results = nullptr;

flux.hpp

@@ -233,14 +233,17 @@ namespace Flux {
     __STATIC_INLINE__ struct ggml_tensor* modulate(struct ggml_context* ctx,
                                                    struct ggml_tensor* x,
                                                    struct ggml_tensor* shift,
-                                                   struct ggml_tensor* scale) {
+                                                   struct ggml_tensor* scale,
+                                                   bool skip_reshape = false) {
         // x: [N, L, C]
         // scale: [N, C]
         // shift: [N, C]
-        scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
-        shift = ggml_reshape_3d(ctx, shift, shift->ne[0], 1, shift->ne[1]);  // [N, 1, C]
-        x     = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
-        x     = ggml_add(ctx, x, shift);
+        if (!skip_reshape) {
+            scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
+            shift = ggml_reshape_3d(ctx, shift, shift->ne[0], 1, shift->ne[1]);  // [N, 1, C]
+        }
+        x = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        x = ggml_add(ctx, x, shift);
         return x;
     }
 

ggml_extend.hpp

@@ -1033,7 +1033,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_pad(struct ggml_context* ctx,
                                                    int p3          = 0,
                                                    bool circular_x = false,
                                                    bool circular_y = false) {
-    return ggml_ext_pad_ext(ctx, x, p0, p0, p1, p1, p2, p2, p3, p3, circular_x, circular_y);
+    return ggml_ext_pad_ext(ctx, x, 0, p0, 0, p1, 0, p2, 0, p3, circular_x, circular_y);
 }
 
 // w: [OC，IC, KH, KW]
@@ -1062,7 +1062,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_conv_2d(struct ggml_context* ctx,
     }
 
     if ((p0 != 0 || p1 != 0) && (circular_x || circular_y)) {
-        x  = ggml_ext_pad(ctx, x, p0, p1, 0, 0, circular_x, circular_y);
+        x  = ggml_ext_pad_ext(ctx, x, p0, p0, p1, p1, 0, 0, 0, 0, circular_x, circular_y);
         p0 = 0;
         p1 = 0;
     }

model.cpp

@@ -1340,7 +1340,7 @@ std::string ModelLoader::load_umt5_tokenizer_json() {
     return json_str;
 }
 
-bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p) {
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
     int64_t process_time_ms = 0;
     std::atomic<int64_t> read_time_ms(0);
     std::atomic<int64_t> memcpy_time_ms(0);
@@ -1390,6 +1390,15 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
             }
         }
 
+        std::unique_ptr<MmapWrapper> mmapped;
+        if (enable_mmap && !is_zip) {
+            LOG_DEBUG("using mmap for I/O");
+            mmapped = MmapWrapper::create(file_path);
+            if (!mmapped) {
+                LOG_WARN("failed to memory-map '%s'", file_path.c_str());
+            }
+        }
+
         int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
         if (n_threads < 1) {
             n_threads = 1;
@@ -1411,7 +1420,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         failed = true;
                         return;
                     }
-                } else {
+                } else if (!mmapped) {
                     file.open(file_path, std::ios::binary);
                     if (!file.is_open()) {
                         LOG_ERROR("failed to open '%s'", file_path.c_str());
@@ -1464,6 +1473,11 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                                 zip_entry_noallocread(zip, (void*)buf, n);
                             }
                             zip_entry_close(zip);
+                        } else if (mmapped) {
+                            if (!mmapped->copy_data(buf, n, tensor_storage.offset)) {
+                                LOG_ERROR("read tensor data failed: '%s'", file_path.c_str());
+                                failed = true;
+                            }
                         } else {
                             file.seekg(tensor_storage.offset);
                             file.read(buf, n);
@@ -1588,7 +1602,8 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
 
 bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
                                std::set<std::string> ignore_tensors,
-                               int n_threads) {
+                               int n_threads,
+                               bool enable_mmap) {
     std::set<std::string> tensor_names_in_file;
     std::mutex tensor_names_mutex;
     auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
@@ -1631,7 +1646,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
         return true;
     };
 
-    bool success = load_tensors(on_new_tensor_cb, n_threads);
+    bool success = load_tensors(on_new_tensor_cb, n_threads, enable_mmap);
     if (!success) {
         LOG_ERROR("load tensors from file failed");
         return false;

model.h

@@ -310,10 +310,11 @@ public:
     std::map<ggml_type, uint32_t> get_vae_wtype_stat();
     String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
     void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0);
+    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0, bool use_mmap = false);
     bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
                       std::set<std::string> ignore_tensors = {},
-                      int n_threads                        = 0);
+                      int n_threads                        = 0,
+                      bool use_mmap                        = false);
 
     std::vector<std::string> get_tensor_names() const {
         std::vector<std::string> names;

name_conversion.cpp

@@ -960,6 +960,7 @@ bool is_first_stage_model_name(const std::string& name) {
 std::string convert_tensor_name(std::string name, SDVersion version) {
     bool is_lora                             = false;
     bool is_lycoris_underline                = false;
+    bool is_underline                        = false;
     std::vector<std::string> lora_prefix_vec = {
         "lora.lora.",
         "lora.lora_",
@@ -967,12 +968,27 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
         "lora.lycoris.",
         "lora.",
     };
+    std::vector<std::string> underline_lora_prefix_vec = {
+        "unet_",
+        "te_",
+        "te1_",
+        "te2_",
+        "te3_",
+        "vae_",
+    };
     for (const auto& prefix : lora_prefix_vec) {
         if (starts_with(name, prefix)) {
             is_lora = true;
             name    = name.substr(prefix.size());
             if (contains(prefix, "lycoris_")) {
                 is_lycoris_underline = true;
+            } else {
+                for (const auto& underline_lora_prefix : underline_lora_prefix_vec) {
+                    if (starts_with(name, underline_lora_prefix)) {
+                        is_underline = true;
+                        break;
+                    }
+                }
             }
             break;
         }
@@ -1034,7 +1050,7 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
 
         // LOG_DEBUG("name %s %d", name.c_str(), version);
 
-        if (sd_version_is_unet(version) || sd_version_is_flux(version) || is_lycoris_underline) {
+        if (sd_version_is_unet(version) || is_underline || is_lycoris_underline) {
             name = convert_sep_to_dot(name);
         }
     }

qwen_image.hpp

@@ -191,11 +191,16 @@ namespace Qwen {
     };
 
     class QwenImageTransformerBlock : public GGMLBlock {
+    protected:
+        bool zero_cond_t;
+
     public:
         QwenImageTransformerBlock(int64_t dim,
                                   int64_t num_attention_heads,
                                   int64_t attention_head_dim,
-                                  float eps = 1e-6) {
+                                  float eps        = 1e-6,
+                                  bool zero_cond_t = false)
+            : zero_cond_t(zero_cond_t) {
             // img_mod.0 is nn.SiLU()
             blocks["img_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
 
@@ -220,11 +225,37 @@ namespace Qwen {
                                                                                eps));
         }
 
+        std::vector<ggml_tensor*> get_mod_params_vec(ggml_context* ctx, ggml_tensor* mod_params, ggml_tensor* index = nullptr) {
+            // index: [N, n_img_token]
+            // mod_params: [N, hidden_size * 12]
+            if (index == nullptr) {
+                return ggml_ext_chunk(ctx, mod_params, 6, 0);
+            }
+            mod_params          = ggml_reshape_1d(ctx, mod_params, ggml_nelements(mod_params));
+            auto mod_params_vec = ggml_ext_chunk(ctx, mod_params, 12, 0);
+            index               = ggml_reshape_3d(ctx, index, 1, index->ne[0], index->ne[1]);                                      // [N, n_img_token, 1]
+            index               = ggml_repeat_4d(ctx, index, mod_params_vec[0]->ne[0], index->ne[1], index->ne[2], index->ne[3]);  // [N, n_img_token, hidden_size]
+            std::vector<ggml_tensor*> mod_results;
+            for (int i = 0; i < 6; i++) {
+                auto mod_0 = mod_params_vec[i];
+                auto mod_1 = mod_params_vec[i + 6];
+
+                // mod_result = torch.where(index == 0, mod_0, mod_1)
+                // mod_result = (1 - index)*mod_0 + index*mod_1
+                mod_0           = ggml_sub(ctx, ggml_repeat(ctx, mod_0, index), ggml_mul(ctx, index, mod_0));  // [N, n_img_token, hidden_size]
+                mod_1           = ggml_mul(ctx, index, mod_1);                                                 // [N, n_img_token, hidden_size]
+                auto mod_result = ggml_add(ctx, mod_0, mod_1);
+                mod_results.push_back(mod_result);
+            }
+            return mod_results;
+        }
+
         virtual std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
                                                               struct ggml_tensor* img,
                                                               struct ggml_tensor* txt,
                                                               struct ggml_tensor* t_emb,
-                                                              struct ggml_tensor* pe) {
+                                                              struct ggml_tensor* pe,
+                                                              struct ggml_tensor* modulate_index = nullptr) {
             // img: [N, n_img_token, hidden_size]
             // txt: [N, n_txt_token, hidden_size]
             // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
@@ -244,14 +275,18 @@ namespace Qwen {
 
             auto img_mod_params    = ggml_silu(ctx->ggml_ctx, t_emb);
             img_mod_params         = img_mod_1->forward(ctx, img_mod_params);
-            auto img_mod_param_vec = ggml_ext_chunk(ctx->ggml_ctx, img_mod_params, 6, 0);
+            auto img_mod_param_vec = get_mod_params_vec(ctx->ggml_ctx, img_mod_params, modulate_index);
+
+            if (zero_cond_t) {
+                t_emb = ggml_ext_chunk(ctx->ggml_ctx, t_emb, 2, 1)[0];
+            }
 
             auto txt_mod_params    = ggml_silu(ctx->ggml_ctx, t_emb);
             txt_mod_params         = txt_mod_1->forward(ctx, txt_mod_params);
-            auto txt_mod_param_vec = ggml_ext_chunk(ctx->ggml_ctx, txt_mod_params, 6, 0);
+            auto txt_mod_param_vec = get_mod_params_vec(ctx->ggml_ctx, txt_mod_params);
 
             auto img_normed    = img_norm1->forward(ctx, img);
-            auto img_modulated = Flux::modulate(ctx->ggml_ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1]);
+            auto img_modulated = Flux::modulate(ctx->ggml_ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1], modulate_index != nullptr);
             auto img_gate1     = img_mod_param_vec[2];
 
             auto txt_normed    = txt_norm1->forward(ctx, txt);
@@ -264,7 +299,7 @@ namespace Qwen {
             txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn_output, txt_gate1));
 
             auto img_normed2    = img_norm2->forward(ctx, img);
-            auto img_modulated2 = Flux::modulate(ctx->ggml_ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4]);
+            auto img_modulated2 = Flux::modulate(ctx->ggml_ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4], modulate_index != nullptr);
             auto img_gate2      = img_mod_param_vec[5];
 
             auto txt_normed2    = txt_norm2->forward(ctx, txt);
@@ -325,6 +360,7 @@ namespace Qwen {
         float theta                 = 10000;
         std::vector<int> axes_dim   = {16, 56, 56};
         int64_t axes_dim_sum        = 128;
+        bool zero_cond_t            = false;
     };
 
     class QwenImageModel : public GGMLBlock {
@@ -346,7 +382,8 @@ namespace Qwen {
                 auto block                                        = std::shared_ptr<GGMLBlock>(new QwenImageTransformerBlock(inner_dim,
                                                                                                                              params.num_attention_heads,
                                                                                                                              params.attention_head_dim,
-                                                                                                                             1e-6f));
+                                                                                                                             1e-6f,
+                                                                                                                             params.zero_cond_t));
                 blocks["transformer_blocks." + std::to_string(i)] = block;
             }
 
@@ -421,7 +458,8 @@ namespace Qwen {
                                          struct ggml_tensor* x,
                                          struct ggml_tensor* timestep,
                                          struct ggml_tensor* context,
-                                         struct ggml_tensor* pe) {
+                                         struct ggml_tensor* pe,
+                                         struct ggml_tensor* modulate_index = nullptr) {
             auto time_text_embed = std::dynamic_pointer_cast<QwenTimestepProjEmbeddings>(blocks["time_text_embed"]);
             auto txt_norm        = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
             auto img_in          = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
@@ -430,18 +468,26 @@ namespace Qwen {
             auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
 
             auto t_emb = time_text_embed->forward(ctx, timestep);
-            auto img   = img_in->forward(ctx, x);
-            auto txt   = txt_norm->forward(ctx, context);
-            txt        = txt_in->forward(ctx, txt);
+            if (params.zero_cond_t) {
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros(ctx->ggml_ctx, timestep->ne[0], timestep->ne[1], timestep->ne[2], timestep->ne[3]));
+                t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
+            }
+            auto img = img_in->forward(ctx, x);
+            auto txt = txt_norm->forward(ctx, context);
+            txt      = txt_in->forward(ctx, txt);
 
             for (int i = 0; i < params.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
 
-                auto result = block->forward(ctx, img, txt, t_emb, pe);
+                auto result = block->forward(ctx, img, txt, t_emb, pe, modulate_index);
                 img         = result.first;
                 txt         = result.second;
             }
 
+            if (params.zero_cond_t) {
+                t_emb = ggml_ext_chunk(ctx->ggml_ctx, t_emb, 2, 1)[0];
+            }
+
             img = norm_out->forward(ctx, img, t_emb);
             img = proj_out->forward(ctx, img);
 
@@ -453,7 +499,8 @@ namespace Qwen {
                                     struct ggml_tensor* timestep,
                                     struct ggml_tensor* context,
                                     struct ggml_tensor* pe,
-                                    std::vector<ggml_tensor*> ref_latents = {}) {
+                                    std::vector<ggml_tensor*> ref_latents = {},
+                                    struct ggml_tensor* modulate_index    = nullptr) {
             // Forward pass of DiT.
             // x: [N, C, H, W]
             // timestep: [N,]
@@ -479,7 +526,7 @@ namespace Qwen {
             int64_t h_len = ((H + (params.patch_size / 2)) / params.patch_size);
             int64_t w_len = ((W + (params.patch_size / 2)) / params.patch_size);
 
-            auto out = forward_orig(ctx, img, timestep, context, pe);  // [N, h_len*w_len, ph*pw*C]
+            auto out = forward_orig(ctx, img, timestep, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]
 
             if (out->ne[1] > img_tokens) {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [num_tokens, N, C * patch_size * patch_size]
@@ -502,19 +549,25 @@ namespace Qwen {
         QwenImageParams qwen_image_params;
         QwenImageModel qwen_image;
         std::vector<float> pe_vec;
+        std::vector<float> modulate_index_vec;
         SDVersion version;
 
         QwenImageRunner(ggml_backend_t backend,
                         bool offload_params_to_cpu,
                         const String2TensorStorage& tensor_storage_map = {},
                         const std::string prefix                       = "",
-                        SDVersion version                              = VERSION_QWEN_IMAGE)
+                        SDVersion version                              = VERSION_QWEN_IMAGE,
+                        bool zero_cond_t                               = false)
             : GGMLRunner(backend, offload_params_to_cpu) {
-            qwen_image_params.num_layers = 0;
+            qwen_image_params.num_layers  = 0;
+            qwen_image_params.zero_cond_t = zero_cond_t;
             for (auto pair : tensor_storage_map) {
                 std::string tensor_name = pair.first;
                 if (tensor_name.find(prefix) == std::string::npos)
                     continue;
+                if (tensor_name.find("__index_timestep_zero__") != std::string::npos) {
+                    qwen_image_params.zero_cond_t = true;
+                }
                 size_t pos = tensor_name.find("transformer_blocks.");
                 if (pos != std::string::npos) {
                     tensor_name = tensor_name.substr(pos);  // remove prefix
@@ -529,6 +582,9 @@ namespace Qwen {
                 }
             }
             LOG_INFO("qwen_image_params.num_layers: %ld", qwen_image_params.num_layers);
+            if (qwen_image_params.zero_cond_t) {
+                LOG_INFO("use zero_cond_t");
+            }
             qwen_image = QwenImageModel(qwen_image_params);
             qwen_image.init(params_ctx, tensor_storage_map, prefix);
         }
@@ -576,6 +632,31 @@ namespace Qwen {
             // pe->data = nullptr;
             set_backend_tensor_data(pe, pe_vec.data());
 
+            ggml_tensor* modulate_index = nullptr;
+            if (qwen_image_params.zero_cond_t) {
+                modulate_index_vec.clear();
+
+                int64_t h_len          = ((x->ne[1] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
+                int64_t w_len          = ((x->ne[0] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
+                int64_t num_img_tokens = h_len * w_len;
+
+                modulate_index_vec.insert(modulate_index_vec.end(), num_img_tokens, 0.f);
+                int64_t num_ref_img_tokens = 0;
+                for (ggml_tensor* ref : ref_latents) {
+                    int64_t h_len = ((ref->ne[1] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
+                    int64_t w_len = ((ref->ne[0] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
+
+                    num_ref_img_tokens += h_len * w_len;
+                }
+
+                if (num_ref_img_tokens > 0) {
+                    modulate_index_vec.insert(modulate_index_vec.end(), num_ref_img_tokens, 1.f);
+                }
+
+                modulate_index = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_F32, modulate_index_vec.size());
+                set_backend_tensor_data(modulate_index, modulate_index_vec.data());
+            }
+
             auto runner_ctx = get_context();
 
             struct ggml_tensor* out = qwen_image.forward(&runner_ctx,
@@ -583,7 +664,8 @@ namespace Qwen {
                                                          timesteps,
                                                          context,
                                                          pe,
-                                                         ref_latents);
+                                                         ref_latents,
+                                                         modulate_index);
 
             ggml_build_forward_expand(gf, out);
 

stable-diffusion.cpp

@@ -7,6 +7,7 @@
 #include "stable-diffusion.h"
 #include "util.h"
 
+#include "cache_dit.hpp"
 #include "conditioner.hpp"
 #include "control.hpp"
 #include "denoiser.hpp"
@@ -16,6 +17,7 @@
 #include "lora.hpp"
 #include "pmid.hpp"
 #include "tae.hpp"
+#include "ucache.hpp"
 #include "vae.hpp"
 
 #include "latent-preview.h"
@@ -127,7 +129,7 @@ public:
     bool use_tiny_autoencoder            = false;
     sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0, 0};
     bool offload_params_to_cpu           = false;
-    bool stacked_id                      = false;
+    bool use_pmid                        = false;
 
     bool is_using_v_parameterization     = false;
     bool is_using_edm_v_parameterization = false;
@@ -518,7 +520,8 @@ public:
                                                                    offload_params_to_cpu,
                                                                    tensor_storage_map,
                                                                    "model.diffusion_model",
-                                                                   version);
+                                                                   version,
+                                                                   sd_ctx_params->qwen_image_zero_cond_t);
             } else if (sd_version_is_z_image(version)) {
                 cond_stage_model = std::make_shared<LLMEmbedder>(clip_backend,
                                                                  offload_params_to_cpu,
@@ -588,8 +591,8 @@ public:
                 vae_backend = backend;
             }
 
-            if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
-                if (!use_tiny_autoencoder) {
+            if (!use_tiny_autoencoder || sd_ctx_params->tae_preview_only) {
+                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
                     first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
                                                                             offload_params_to_cpu,
                                                                             tensor_storage_map,
@@ -598,57 +601,56 @@ public:
                                                                             version);
                     first_stage_model->alloc_params_buffer();
                     first_stage_model->get_param_tensors(tensors, "first_stage_model");
+                } else if (version == VERSION_CHROMA_RADIANCE) {
+                    first_stage_model = std::make_shared<FakeVAE>(vae_backend,
+                                                                  offload_params_to_cpu);
                 } else {
+                    first_stage_model = std::make_shared<AutoEncoderKL>(vae_backend,
+                                                                        offload_params_to_cpu,
+                                                                        tensor_storage_map,
+                                                                        "first_stage_model",
+                                                                        vae_decode_only,
+                                                                        false,
+                                                                        version);
+                    if (sd_ctx_params->vae_conv_direct) {
+                        LOG_INFO("Using Conv2d direct in the vae model");
+                        first_stage_model->set_conv2d_direct_enabled(true);
+                    }
+                    if (version == VERSION_SDXL &&
+                        (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale)) {
+                        float vae_conv_2d_scale = 1.f / 32.f;
+                        LOG_WARN(
+                            "No VAE specified with --vae or --force-sdxl-vae-conv-scale flag set, "
+                            "using Conv2D scale %.3f",
+                            vae_conv_2d_scale);
+                        first_stage_model->set_conv2d_scale(vae_conv_2d_scale);
+                    }
+                    first_stage_model->alloc_params_buffer();
+                    first_stage_model->get_param_tensors(tensors, "first_stage_model");
+                }
+            }
+
+            if (use_tiny_autoencoder) {
+                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
                     tae_first_stage = std::make_shared<TinyVideoAutoEncoder>(vae_backend,
                                                                              offload_params_to_cpu,
                                                                              tensor_storage_map,
                                                                              "decoder",
                                                                              vae_decode_only,
                                                                              version);
-                    if (sd_ctx_params->vae_conv_direct) {
-                        LOG_INFO("Using Conv2d direct in the tae model");
-                        tae_first_stage->set_conv2d_direct_enabled(true);
-                    }
+                } else {
+                    tae_first_stage = std::make_shared<TinyImageAutoEncoder>(vae_backend,
+                                                                             offload_params_to_cpu,
+                                                                             tensor_storage_map,
+                                                                             "decoder.layers",
+                                                                             vae_decode_only,
+                                                                             version);
                 }
-            } else if (version == VERSION_CHROMA_RADIANCE) {
-                first_stage_model = std::make_shared<FakeVAE>(vae_backend,
-                                                              offload_params_to_cpu);
-            } else if (!use_tiny_autoencoder || sd_ctx_params->tae_preview_only) {
-                first_stage_model = std::make_shared<AutoEncoderKL>(vae_backend,
-                                                                    offload_params_to_cpu,
-                                                                    tensor_storage_map,
-                                                                    "first_stage_model",
-                                                                    vae_decode_only,
-                                                                    false,
-                                                                    version);
-                if (sd_ctx_params->vae_conv_direct) {
-                    LOG_INFO("Using Conv2d direct in the vae model");
-                    first_stage_model->set_conv2d_direct_enabled(true);
-                }
-                if (version == VERSION_SDXL &&
-                    (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale)) {
-                    float vae_conv_2d_scale = 1.f / 32.f;
-                    LOG_WARN(
-                        "No VAE specified with --vae or --force-sdxl-vae-conv-scale flag set, "
-                        "using Conv2D scale %.3f",
-                        vae_conv_2d_scale);
-                    first_stage_model->set_conv2d_scale(vae_conv_2d_scale);
-                }
-                first_stage_model->alloc_params_buffer();
-                first_stage_model->get_param_tensors(tensors, "first_stage_model");
-            } else if (use_tiny_autoencoder) {
-                tae_first_stage = std::make_shared<TinyImageAutoEncoder>(vae_backend,
-                                                                         offload_params_to_cpu,
-                                                                         tensor_storage_map,
-                                                                         "decoder.layers",
-                                                                         vae_decode_only,
-                                                                         version);
                 if (sd_ctx_params->vae_conv_direct) {
                     LOG_INFO("Using Conv2d direct in the tae model");
                     tae_first_stage->set_conv2d_direct_enabled(true);
                 }
             }
-            // first_stage_model->get_param_tensors(tensors, "first_stage_model.");
 
             if (strlen(SAFE_STR(sd_ctx_params->control_net_path)) > 0) {
                 ggml_backend_t controlnet_backend = nullptr;
@@ -699,10 +701,10 @@ public:
                 if (!model_loader.init_from_file_and_convert_name(sd_ctx_params->photo_maker_path, "pmid.")) {
                     LOG_WARN("loading stacked ID embedding from '%s' failed", sd_ctx_params->photo_maker_path);
                 } else {
-                    stacked_id = true;
+                    use_pmid = true;
                 }
             }
-            if (stacked_id) {
+            if (use_pmid) {
                 if (!pmid_model->alloc_params_buffer()) {
                     LOG_ERROR(" pmid model params buffer allocation failed");
                     return false;
@@ -743,11 +745,12 @@ public:
         if (use_tiny_autoencoder) {
             ignore_tensors.insert("first_stage_model.");
         }
-        if (stacked_id) {
+        if (use_pmid) {
             ignore_tensors.insert("pmid.unet.");
         }
         ignore_tensors.insert("model.diffusion_model.__x0__");
         ignore_tensors.insert("model.diffusion_model.__32x32__");
+        ignore_tensors.insert("model.diffusion_model.__index_timestep_zero__");
 
         if (vae_decode_only) {
             ignore_tensors.insert("first_stage_model.encoder");
@@ -763,7 +766,7 @@ public:
         if (version == VERSION_SVD) {
             ignore_tensors.insert("conditioner.embedders.3");
         }
-        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads);
+        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads, sd_ctx_params->enable_mmap);
         if (!success) {
             LOG_ERROR("load tensors from model loader failed");
             ggml_free(ctx);
@@ -796,7 +799,7 @@ public:
                 control_net_params_mem_size = control_net->get_params_buffer_size();
             }
             size_t pmid_params_mem_size = 0;
-            if (stacked_id) {
+            if (use_pmid) {
                 pmid_params_mem_size = pmid_model->get_params_buffer_size();
             }
 
@@ -1208,14 +1211,89 @@ public:
         }
     }
 
-    ggml_tensor* id_encoder(ggml_context* work_ctx,
-                            ggml_tensor* init_img,
-                            ggml_tensor* prompts_embeds,
-                            ggml_tensor* id_embeds,
-                            std::vector<bool>& class_tokens_mask) {
-        ggml_tensor* res = nullptr;
-        pmid_model->compute(n_threads, init_img, prompts_embeds, id_embeds, class_tokens_mask, &res, work_ctx);
-        return res;
+    SDCondition get_pmid_conditon(ggml_context* work_ctx,
+                                  sd_pm_params_t pm_params,
+                                  ConditionerParams& condition_params) {
+        SDCondition id_cond;
+        if (use_pmid) {
+            if (!pmid_lora->applied) {
+                int64_t t0 = ggml_time_ms();
+                pmid_lora->apply(tensors, version, n_threads);
+                int64_t t1         = ggml_time_ms();
+                pmid_lora->applied = true;
+                LOG_INFO("pmid_lora apply completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+                if (free_params_immediately) {
+                    pmid_lora->free_params_buffer();
+                }
+            }
+            // preprocess input id images
+            bool pmv2 = pmid_model->get_version() == PM_VERSION_2;
+            if (pm_params.id_images_count > 0) {
+                int clip_image_size        = 224;
+                pmid_model->style_strength = pm_params.style_strength;
+
+                auto id_image_tensor = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, clip_image_size, clip_image_size, 3, pm_params.id_images_count);
+
+                std::vector<sd_image_f32_t> processed_id_images;
+                for (int i = 0; i < pm_params.id_images_count; i++) {
+                    sd_image_f32_t id_image           = sd_image_t_to_sd_image_f32_t(pm_params.id_images[i]);
+                    sd_image_f32_t processed_id_image = clip_preprocess(id_image, clip_image_size, clip_image_size);
+                    free(id_image.data);
+                    id_image.data = nullptr;
+                    processed_id_images.push_back(processed_id_image);
+                }
+
+                ggml_ext_tensor_iter(id_image_tensor, [&](ggml_tensor* id_image_tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+                    float value = sd_image_get_f32(processed_id_images[i3], i0, i1, i2, false);
+                    ggml_ext_tensor_set_f32(id_image_tensor, value, i0, i1, i2, i3);
+                });
+
+                for (auto& image : processed_id_images) {
+                    free(image.data);
+                    image.data = nullptr;
+                }
+                processed_id_images.clear();
+
+                int64_t t0                      = ggml_time_ms();
+                condition_params.num_input_imgs = pm_params.id_images_count;
+                auto cond_tup                   = cond_stage_model->get_learned_condition_with_trigger(work_ctx,
+                                                                                                       n_threads,
+                                                                                                       condition_params);
+                id_cond                         = std::get<0>(cond_tup);
+                auto class_tokens_mask          = std::get<1>(cond_tup);
+                struct ggml_tensor* id_embeds   = nullptr;
+                if (pmv2 && pm_params.id_embed_path != nullptr) {
+                    id_embeds = load_tensor_from_file(work_ctx, pm_params.id_embed_path);
+                }
+                if (pmv2 && id_embeds == nullptr) {
+                    LOG_WARN("Provided PhotoMaker images, but NO valid ID embeds file for PM v2");
+                    LOG_WARN("Turn off PhotoMaker");
+                    use_pmid = false;
+                } else {
+                    if (pmv2 && pm_params.id_images_count != id_embeds->ne[1]) {
+                        LOG_WARN("PhotoMaker image count (%d) does NOT match ID embeds (%d). You should run face_detect.py again.", pm_params.id_images_count, id_embeds->ne[1]);
+                        LOG_WARN("Turn off PhotoMaker");
+                        use_pmid = false;
+                    } else {
+                        ggml_tensor* res = nullptr;
+                        pmid_model->compute(n_threads, id_image_tensor, id_cond.c_crossattn, id_embeds, class_tokens_mask, &res, work_ctx);
+                        id_cond.c_crossattn = res;
+                        int64_t t1          = ggml_time_ms();
+                        LOG_INFO("Photomaker ID Stacking, taking %" PRId64 " ms", t1 - t0);
+                        if (free_params_immediately) {
+                            pmid_model->free_params_buffer();
+                        }
+                        // Encode input prompt without the trigger word for delayed conditioning
+                        condition_params.text = cond_stage_model->remove_trigger_from_prompt(work_ctx, condition_params.text);
+                    }
+                }
+            } else {
+                LOG_WARN("Provided PhotoMaker model file, but NO input ID images");
+                LOG_WARN("Turn off PhotoMaker");
+                use_pmid = false;
+            }
+        }
+        return id_cond;
     }
 
     ggml_tensor* get_clip_vision_output(ggml_context* work_ctx,
@@ -1525,12 +1603,12 @@ public:
                         const std::vector<float>& sigmas,
                         int start_merge_step,
                         SDCondition id_cond,
-                        std::vector<ggml_tensor*> ref_latents         = {},
-                        bool increase_ref_index                       = false,
-                        ggml_tensor* denoise_mask                     = nullptr,
-                        ggml_tensor* vace_context                     = nullptr,
-                        float vace_strength                           = 1.f,
-                        const sd_easycache_params_t* easycache_params = nullptr) {
+                        std::vector<ggml_tensor*> ref_latents = {},
+                        bool increase_ref_index               = false,
+                        ggml_tensor* denoise_mask             = nullptr,
+                        ggml_tensor* vace_context             = nullptr,
+                        float vace_strength                   = 1.f,
+                        const sd_cache_params_t* cache_params = nullptr) {
         if (shifted_timestep > 0 && !sd_version_is_sdxl(version)) {
             LOG_WARN("timestep shifting is only supported for SDXL models!");
             shifted_timestep = 0;
@@ -1558,31 +1636,40 @@ public:
         }
 
         EasyCacheState easycache_state;
+        UCacheState ucache_state;
+        CacheDitConditionState cachedit_state;
         bool easycache_enabled = false;
-        if (easycache_params != nullptr && easycache_params->enabled) {
-            bool easycache_supported = sd_version_is_dit(version);
-            if (!easycache_supported) {
-                LOG_WARN("EasyCache requested but not supported for this model type");
-            } else {
-                EasyCacheConfig easycache_config;
-                easycache_config.enabled         = true;
-                easycache_config.reuse_threshold = std::max(0.0f, easycache_params->reuse_threshold);
-                easycache_config.start_percent   = easycache_params->start_percent;
-                easycache_config.end_percent     = easycache_params->end_percent;
-                bool percent_valid               = easycache_config.start_percent >= 0.0f &&
-                                     easycache_config.start_percent < 1.0f &&
-                                     easycache_config.end_percent > 0.0f &&
-                                     easycache_config.end_percent <= 1.0f &&
-                                     easycache_config.start_percent < easycache_config.end_percent;
-                if (!percent_valid) {
-                    LOG_WARN("EasyCache disabled due to invalid percent range (start=%.3f, end=%.3f)",
-                             easycache_config.start_percent,
-                             easycache_config.end_percent);
+        bool ucache_enabled    = false;
+        bool cachedit_enabled  = false;
+
+        if (cache_params != nullptr && cache_params->mode != SD_CACHE_DISABLED) {
+            bool percent_valid = true;
+            if (cache_params->mode == SD_CACHE_EASYCACHE || cache_params->mode == SD_CACHE_UCACHE) {
+                percent_valid = cache_params->start_percent >= 0.0f &&
+                                cache_params->start_percent < 1.0f &&
+                                cache_params->end_percent > 0.0f &&
+                                cache_params->end_percent <= 1.0f &&
+                                cache_params->start_percent < cache_params->end_percent;
+            }
+
+            if (!percent_valid) {
+                LOG_WARN("Cache disabled due to invalid percent range (start=%.3f, end=%.3f)",
+                         cache_params->start_percent,
+                         cache_params->end_percent);
+            } else if (cache_params->mode == SD_CACHE_EASYCACHE) {
+                bool easycache_supported = sd_version_is_dit(version);
+                if (!easycache_supported) {
+                    LOG_WARN("EasyCache requested but not supported for this model type");
                 } else {
+                    EasyCacheConfig easycache_config;
+                    easycache_config.enabled         = true;
+                    easycache_config.reuse_threshold = std::max(0.0f, cache_params->reuse_threshold);
+                    easycache_config.start_percent   = cache_params->start_percent;
+                    easycache_config.end_percent     = cache_params->end_percent;
                     easycache_state.init(easycache_config, denoiser.get());
                     if (easycache_state.enabled()) {
                         easycache_enabled = true;
-                        LOG_INFO("EasyCache enabled - threshold: %.3f, start_percent: %.2f, end_percent: %.2f",
+                        LOG_INFO("EasyCache enabled - threshold: %.3f, start: %.2f, end: %.2f",
                                  easycache_config.reuse_threshold,
                                  easycache_config.start_percent,
                                  easycache_config.end_percent);
@@ -1590,9 +1677,84 @@ public:
                         LOG_WARN("EasyCache requested but could not be initialized for this run");
                     }
                 }
+            } else if (cache_params->mode == SD_CACHE_UCACHE) {
+                bool ucache_supported = sd_version_is_unet(version);
+                if (!ucache_supported) {
+                    LOG_WARN("UCache requested but not supported for this model type (only UNET models)");
+                } else {
+                    UCacheConfig ucache_config;
+                    ucache_config.enabled                = true;
+                    ucache_config.reuse_threshold        = std::max(0.0f, cache_params->reuse_threshold);
+                    ucache_config.start_percent          = cache_params->start_percent;
+                    ucache_config.end_percent            = cache_params->end_percent;
+                    ucache_config.error_decay_rate       = std::max(0.0f, std::min(1.0f, cache_params->error_decay_rate));
+                    ucache_config.use_relative_threshold = cache_params->use_relative_threshold;
+                    ucache_config.reset_error_on_compute = cache_params->reset_error_on_compute;
+                    ucache_state.init(ucache_config, denoiser.get());
+                    if (ucache_state.enabled()) {
+                        ucache_enabled = true;
+                        LOG_INFO("UCache enabled - threshold: %.3f, start: %.2f, end: %.2f, decay: %.2f, relative: %s, reset: %s",
+                                 ucache_config.reuse_threshold,
+                                 ucache_config.start_percent,
+                                 ucache_config.end_percent,
+                                 ucache_config.error_decay_rate,
+                                 ucache_config.use_relative_threshold ? "true" : "false",
+                                 ucache_config.reset_error_on_compute ? "true" : "false");
+                    } else {
+                        LOG_WARN("UCache requested but could not be initialized for this run");
+                    }
+                }
+            } else if (cache_params->mode == SD_CACHE_DBCACHE ||
+                       cache_params->mode == SD_CACHE_TAYLORSEER ||
+                       cache_params->mode == SD_CACHE_CACHE_DIT) {
+                bool cachedit_supported = sd_version_is_dit(version);
+                if (!cachedit_supported) {
+                    LOG_WARN("CacheDIT requested but not supported for this model type (only DiT models)");
+                } else {
+                    DBCacheConfig dbcfg;
+                    dbcfg.enabled                     = (cache_params->mode == SD_CACHE_DBCACHE ||
+                                     cache_params->mode == SD_CACHE_CACHE_DIT);
+                    dbcfg.Fn_compute_blocks           = cache_params->Fn_compute_blocks;
+                    dbcfg.Bn_compute_blocks           = cache_params->Bn_compute_blocks;
+                    dbcfg.residual_diff_threshold     = cache_params->residual_diff_threshold;
+                    dbcfg.max_warmup_steps            = cache_params->max_warmup_steps;
+                    dbcfg.max_cached_steps            = cache_params->max_cached_steps;
+                    dbcfg.max_continuous_cached_steps = cache_params->max_continuous_cached_steps;
+                    if (cache_params->scm_mask != nullptr && strlen(cache_params->scm_mask) > 0) {
+                        dbcfg.steps_computation_mask = parse_scm_mask(cache_params->scm_mask);
+                    }
+                    dbcfg.scm_policy_dynamic = cache_params->scm_policy_dynamic;
+
+                    TaylorSeerConfig tcfg;
+                    tcfg.enabled             = (cache_params->mode == SD_CACHE_TAYLORSEER ||
+                                    cache_params->mode == SD_CACHE_CACHE_DIT);
+                    tcfg.n_derivatives       = cache_params->taylorseer_n_derivatives;
+                    tcfg.skip_interval_steps = cache_params->taylorseer_skip_interval;
+
+                    cachedit_state.init(dbcfg, tcfg);
+                    if (cachedit_state.enabled()) {
+                        cachedit_enabled = true;
+                        LOG_INFO("CacheDIT enabled - mode: %s, Fn: %d, Bn: %d, threshold: %.3f, warmup: %d",
+                                 cache_params->mode == SD_CACHE_CACHE_DIT ? "DBCache+TaylorSeer" : (cache_params->mode == SD_CACHE_DBCACHE ? "DBCache" : "TaylorSeer"),
+                                 dbcfg.Fn_compute_blocks,
+                                 dbcfg.Bn_compute_blocks,
+                                 dbcfg.residual_diff_threshold,
+                                 dbcfg.max_warmup_steps);
+                    } else {
+                        LOG_WARN("CacheDIT requested but could not be initialized for this run");
+                    }
+                }
             }
         }
 
+        if (ucache_enabled) {
+            ucache_state.set_sigmas(sigmas);
+        }
+
+        if (cachedit_enabled) {
+            cachedit_state.set_sigmas(sigmas);
+        }
+
         size_t steps          = sigmas.size() - 1;
         struct ggml_tensor* x = ggml_dup_tensor(work_ctx, init_latent);
         copy_ggml_tensor(x, init_latent);
@@ -1696,6 +1858,91 @@ public:
                 return easycache_step_active && easycache_state.is_step_skipped();
             };
 
+            const bool ucache_step_active = ucache_enabled && step > 0;
+            int ucache_step_index         = ucache_step_active ? (step - 1) : -1;
+            if (ucache_step_active) {
+                ucache_state.begin_step(ucache_step_index, sigma);
+            }
+
+            auto ucache_before_condition = [&](const SDCondition* condition, struct ggml_tensor* output_tensor) -> bool {
+                if (!ucache_step_active || condition == nullptr || output_tensor == nullptr) {
+                    return false;
+                }
+                return ucache_state.before_condition(condition,
+                                                     diffusion_params.x,
+                                                     output_tensor,
+                                                     sigma,
+                                                     ucache_step_index);
+            };
+
+            auto ucache_after_condition = [&](const SDCondition* condition, struct ggml_tensor* output_tensor) {
+                if (!ucache_step_active || condition == nullptr || output_tensor == nullptr) {
+                    return;
+                }
+                ucache_state.after_condition(condition,
+                                             diffusion_params.x,
+                                             output_tensor);
+            };
+
+            auto ucache_step_is_skipped = [&]() {
+                return ucache_step_active && ucache_state.is_step_skipped();
+            };
+
+            const bool cachedit_step_active = cachedit_enabled && step > 0;
+            int cachedit_step_index         = cachedit_step_active ? (step - 1) : -1;
+            if (cachedit_step_active) {
+                cachedit_state.begin_step(cachedit_step_index, sigma);
+            }
+
+            auto cachedit_before_condition = [&](const SDCondition* condition, struct ggml_tensor* output_tensor) -> bool {
+                if (!cachedit_step_active || condition == nullptr || output_tensor == nullptr) {
+                    return false;
+                }
+                return cachedit_state.before_condition(condition,
+                                                       diffusion_params.x,
+                                                       output_tensor,
+                                                       sigma,
+                                                       cachedit_step_index);
+            };
+
+            auto cachedit_after_condition = [&](const SDCondition* condition, struct ggml_tensor* output_tensor) {
+                if (!cachedit_step_active || condition == nullptr || output_tensor == nullptr) {
+                    return;
+                }
+                cachedit_state.after_condition(condition,
+                                               diffusion_params.x,
+                                               output_tensor);
+            };
+
+            auto cachedit_step_is_skipped = [&]() {
+                return cachedit_step_active && cachedit_state.is_step_skipped();
+            };
+
+            auto cache_before_condition = [&](const SDCondition* condition, struct ggml_tensor* output_tensor) -> bool {
+                if (easycache_step_active) {
+                    return easycache_before_condition(condition, output_tensor);
+                } else if (ucache_step_active) {
+                    return ucache_before_condition(condition, output_tensor);
+                } else if (cachedit_step_active) {
+                    return cachedit_before_condition(condition, output_tensor);
+                }
+                return false;
+            };
+
+            auto cache_after_condition = [&](const SDCondition* condition, struct ggml_tensor* output_tensor) {
+                if (easycache_step_active) {
+                    easycache_after_condition(condition, output_tensor);
+                } else if (ucache_step_active) {
+                    ucache_after_condition(condition, output_tensor);
+                } else if (cachedit_step_active) {
+                    cachedit_after_condition(condition, output_tensor);
+                }
+            };
+
+            auto cache_step_is_skipped = [&]() {
+                return easycache_step_is_skipped() || ucache_step_is_skipped() || cachedit_step_is_skipped();
+            };
+
             std::vector<float> scaling = denoiser->get_scalings(sigma);
             GGML_ASSERT(scaling.size() == 3);
             float c_skip = scaling[0];
@@ -1771,7 +2018,7 @@ public:
                 active_condition          = &id_cond;
             }
 
-            bool skip_model = easycache_before_condition(active_condition, *active_output);
+            bool skip_model = cache_before_condition(active_condition, *active_output);
             if (!skip_model) {
                 if (!work_diffusion_model->compute(n_threads,
                                                    diffusion_params,
@@ -1779,10 +2026,10 @@ public:
                     LOG_ERROR("diffusion model compute failed");
                     return nullptr;
                 }
-                easycache_after_condition(active_condition, *active_output);
+                cache_after_condition(active_condition, *active_output);
             }
 
-            bool current_step_skipped = easycache_step_is_skipped();
+            bool current_step_skipped = cache_step_is_skipped();
 
             float* negative_data = nullptr;
             if (has_unconditioned) {
@@ -1794,12 +2041,12 @@ public:
                         LOG_ERROR("controlnet compute failed");
                     }
                 }
-                current_step_skipped      = easycache_step_is_skipped();
+                current_step_skipped      = cache_step_is_skipped();
                 diffusion_params.controls = controls;
                 diffusion_params.context  = uncond.c_crossattn;
                 diffusion_params.c_concat = uncond.c_concat;
                 diffusion_params.y        = uncond.c_vector;
-                bool skip_uncond          = easycache_before_condition(&uncond, out_uncond);
+                bool skip_uncond          = cache_before_condition(&uncond, out_uncond);
                 if (!skip_uncond) {
                     if (!work_diffusion_model->compute(n_threads,
                                                        diffusion_params,
@@ -1807,7 +2054,7 @@ public:
                         LOG_ERROR("diffusion model compute failed");
                         return nullptr;
                     }
-                    easycache_after_condition(&uncond, out_uncond);
+                    cache_after_condition(&uncond, out_uncond);
                 }
                 negative_data = (float*)out_uncond->data;
             }
@@ -1817,7 +2064,7 @@ public:
                 diffusion_params.context  = img_cond.c_crossattn;
                 diffusion_params.c_concat = img_cond.c_concat;
                 diffusion_params.y        = img_cond.c_vector;
-                bool skip_img_cond        = easycache_before_condition(&img_cond, out_img_cond);
+                bool skip_img_cond        = cache_before_condition(&img_cond, out_img_cond);
                 if (!skip_img_cond) {
                     if (!work_diffusion_model->compute(n_threads,
                                                        diffusion_params,
@@ -1825,7 +2072,7 @@ public:
                         LOG_ERROR("diffusion model compute failed");
                         return nullptr;
                     }
-                    easycache_after_condition(&img_cond, out_img_cond);
+                    cache_after_condition(&img_cond, out_img_cond);
                 }
                 img_cond_data = (float*)out_img_cond->data;
             }
@@ -1835,7 +2082,7 @@ public:
             float* skip_layer_data = has_skiplayer ? (float*)out_skip->data : nullptr;
             if (is_skiplayer_step) {
                 LOG_DEBUG("Skipping layers at step %d\n", step);
-                if (!easycache_step_is_skipped()) {
+                if (!cache_step_is_skipped()) {
                     // skip layer (same as conditioned)
                     diffusion_params.context     = cond.c_crossattn;
                     diffusion_params.c_concat    = cond.c_concat;
@@ -1939,6 +2186,48 @@ public:
             }
         }
 
+        if (ucache_enabled) {
+            size_t total_steps = sigmas.size() > 0 ? sigmas.size() - 1 : 0;
+            if (ucache_state.total_steps_skipped > 0 && total_steps > 0) {
+                if (ucache_state.total_steps_skipped < static_cast<int>(total_steps)) {
+                    double speedup = static_cast<double>(total_steps) /
+                                     static_cast<double>(total_steps - ucache_state.total_steps_skipped);
+                    LOG_INFO("UCache skipped %d/%zu steps (%.2fx estimated speedup)",
+                             ucache_state.total_steps_skipped,
+                             total_steps,
+                             speedup);
+                } else {
+                    LOG_INFO("UCache skipped %d/%zu steps",
+                             ucache_state.total_steps_skipped,
+                             total_steps);
+                }
+            } else if (total_steps > 0) {
+                LOG_INFO("UCache completed without skipping steps");
+            }
+        }
+
+        if (cachedit_enabled) {
+            size_t total_steps = sigmas.size() > 0 ? sigmas.size() - 1 : 0;
+            if (cachedit_state.total_steps_skipped > 0 && total_steps > 0) {
+                if (cachedit_state.total_steps_skipped < static_cast<int>(total_steps)) {
+                    double speedup = static_cast<double>(total_steps) /
+                                     static_cast<double>(total_steps - cachedit_state.total_steps_skipped);
+                    LOG_INFO("CacheDIT skipped %d/%zu steps (%.2fx estimated speedup), accum_diff: %.4f",
+                             cachedit_state.total_steps_skipped,
+                             total_steps,
+                             speedup,
+                             cachedit_state.accumulated_residual_diff);
+                } else {
+                    LOG_INFO("CacheDIT skipped %d/%zu steps, accum_diff: %.4f",
+                             cachedit_state.total_steps_skipped,
+                             total_steps,
+                             cachedit_state.accumulated_residual_diff);
+                }
+            } else if (total_steps > 0) {
+                LOG_INFO("CacheDIT completed without skipping steps");
+            }
+        }
+
         if (inverse_noise_scaling) {
             x = denoiser->inverse_noise_scaling(sigmas[sigmas.size() - 1], x);
         }
@@ -2554,12 +2843,25 @@ enum lora_apply_mode_t str_to_lora_apply_mode(const char* str) {
     return LORA_APPLY_MODE_COUNT;
 }
 
-void sd_easycache_params_init(sd_easycache_params_t* easycache_params) {
-    *easycache_params                 = {};
-    easycache_params->enabled         = false;
-    easycache_params->reuse_threshold = 0.2f;
-    easycache_params->start_percent   = 0.15f;
-    easycache_params->end_percent     = 0.95f;
+void sd_cache_params_init(sd_cache_params_t* cache_params) {
+    *cache_params                             = {};
+    cache_params->mode                        = SD_CACHE_DISABLED;
+    cache_params->reuse_threshold             = 1.0f;
+    cache_params->start_percent               = 0.15f;
+    cache_params->end_percent                 = 0.95f;
+    cache_params->error_decay_rate            = 1.0f;
+    cache_params->use_relative_threshold      = true;
+    cache_params->reset_error_on_compute      = true;
+    cache_params->Fn_compute_blocks           = 8;
+    cache_params->Bn_compute_blocks           = 0;
+    cache_params->residual_diff_threshold     = 0.08f;
+    cache_params->max_warmup_steps            = 8;
+    cache_params->max_cached_steps            = -1;
+    cache_params->max_continuous_cached_steps = -1;
+    cache_params->taylorseer_n_derivatives    = 1;
+    cache_params->taylorseer_skip_interval    = 1;
+    cache_params->scm_mask                    = nullptr;
+    cache_params->scm_policy_dynamic          = true;
 }
 
 void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
@@ -2573,6 +2875,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->prediction              = PREDICTION_COUNT;
     sd_ctx_params->lora_apply_mode         = LORA_APPLY_AUTO;
     sd_ctx_params->offload_params_to_cpu   = false;
+    sd_ctx_params->enable_mmap             = false;
     sd_ctx_params->keep_clip_on_cpu        = false;
     sd_ctx_params->keep_control_net_on_cpu = false;
     sd_ctx_params->keep_vae_on_cpu         = false;
@@ -2724,7 +3027,7 @@ void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params) {
     sd_img_gen_params->control_strength  = 0.9f;
     sd_img_gen_params->pm_params         = {nullptr, 0, nullptr, 20.f};
     sd_img_gen_params->vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
-    sd_easycache_params_init(&sd_img_gen_params->easycache);
+    sd_cache_params_init(&sd_img_gen_params->cache);
 }
 
 char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
@@ -2768,12 +3071,18 @@ char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
              sd_img_gen_params->pm_params.id_images_count,
              SAFE_STR(sd_img_gen_params->pm_params.id_embed_path),
              BOOL_STR(sd_img_gen_params->vae_tiling_params.enabled));
+    const char* cache_mode_str = "disabled";
+    if (sd_img_gen_params->cache.mode == SD_CACHE_EASYCACHE) {
+        cache_mode_str = "easycache";
+    } else if (sd_img_gen_params->cache.mode == SD_CACHE_UCACHE) {
+        cache_mode_str = "ucache";
+    }
     snprintf(buf + strlen(buf), 4096 - strlen(buf),
-             "easycache: %s (threshold=%.3f, start=%.2f, end=%.2f)\n",
-             sd_img_gen_params->easycache.enabled ? "enabled" : "disabled",
-             sd_img_gen_params->easycache.reuse_threshold,
-             sd_img_gen_params->easycache.start_percent,
-             sd_img_gen_params->easycache.end_percent);
+             "cache: %s (threshold=%.3f, start=%.2f, end=%.2f)\n",
+             cache_mode_str,
+             sd_img_gen_params->cache.reuse_threshold,
+             sd_img_gen_params->cache.start_percent,
+             sd_img_gen_params->cache.end_percent);
     free(sample_params_str);
     return buf;
 }
@@ -2790,7 +3099,7 @@ void sd_vid_gen_params_init(sd_vid_gen_params_t* sd_vid_gen_params) {
     sd_vid_gen_params->video_frames                          = 6;
     sd_vid_gen_params->moe_boundary                          = 0.875f;
     sd_vid_gen_params->vace_strength                         = 1.f;
-    sd_easycache_params_init(&sd_vid_gen_params->easycache);
+    sd_cache_params_init(&sd_vid_gen_params->cache);
 }
 
 struct sd_ctx_t {
@@ -2869,9 +3178,9 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
                                     std::vector<sd_image_t*> ref_images,
                                     std::vector<ggml_tensor*> ref_latents,
                                     bool increase_ref_index,
-                                    ggml_tensor* concat_latent                    = nullptr,
-                                    ggml_tensor* denoise_mask                     = nullptr,
-                                    const sd_easycache_params_t* easycache_params = nullptr) {
+                                    ggml_tensor* concat_latent            = nullptr,
+                                    ggml_tensor* denoise_mask             = nullptr,
+                                    const sd_cache_params_t* cache_params = nullptr) {
     if (seed < 0) {
         // Generally, when using the provided command line, the seed is always >0.
         // However, to prevent potential issues if 'stable-diffusion.cpp' is invoked as a library
@@ -2884,114 +3193,22 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
         guidance.img_cfg = guidance.txt_cfg;
     }
 
-    // for (auto v : sigmas) {
-    //     std::cout << v << " ";
-    // }
-    // std::cout << std::endl;
-
     int sample_steps = sigmas.size() - 1;
 
     int64_t t0 = ggml_time_ms();
 
-    // Photo Maker
-    std::string prompt_text_only;
-    ggml_tensor* init_img = nullptr;
-    SDCondition id_cond;
-    std::vector<bool> class_tokens_mask;
-
     ConditionerParams condition_params;
+    condition_params.text            = prompt;
     condition_params.clip_skip       = clip_skip;
     condition_params.width           = width;
     condition_params.height          = height;
     condition_params.ref_images      = ref_images;
     condition_params.adm_in_channels = sd_ctx->sd->diffusion_model->get_adm_in_channels();
 
-    if (sd_ctx->sd->stacked_id) {
-        if (!sd_ctx->sd->pmid_lora->applied) {
-            int64_t t0 = ggml_time_ms();
-            sd_ctx->sd->pmid_lora->apply(sd_ctx->sd->tensors, sd_ctx->sd->version, sd_ctx->sd->n_threads);
-            int64_t t1                     = ggml_time_ms();
-            sd_ctx->sd->pmid_lora->applied = true;
-            LOG_INFO("pmid_lora apply completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
-            if (sd_ctx->sd->free_params_immediately) {
-                sd_ctx->sd->pmid_lora->free_params_buffer();
-            }
-        }
-        // preprocess input id images
-        bool pmv2 = sd_ctx->sd->pmid_model->get_version() == PM_VERSION_2;
-        if (pm_params.id_images_count > 0) {
-            int clip_image_size                    = 224;
-            sd_ctx->sd->pmid_model->style_strength = pm_params.style_strength;
-
-            init_img = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, clip_image_size, clip_image_size, 3, pm_params.id_images_count);
-
-            std::vector<sd_image_f32_t> processed_id_images;
-            for (int i = 0; i < pm_params.id_images_count; i++) {
-                sd_image_f32_t id_image           = sd_image_t_to_sd_image_f32_t(pm_params.id_images[i]);
-                sd_image_f32_t processed_id_image = clip_preprocess(id_image, clip_image_size, clip_image_size);
-                free(id_image.data);
-                id_image.data = nullptr;
-                processed_id_images.push_back(processed_id_image);
-            }
-
-            ggml_ext_tensor_iter(init_img, [&](ggml_tensor* init_img, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
-                float value = sd_image_get_f32(processed_id_images[i3], i0, i1, i2, false);
-                ggml_ext_tensor_set_f32(init_img, value, i0, i1, i2, i3);
-            });
-
-            for (auto& image : processed_id_images) {
-                free(image.data);
-                image.data = nullptr;
-            }
-            processed_id_images.clear();
-
-            int64_t t0                      = ggml_time_ms();
-            condition_params.text           = prompt;
-            condition_params.num_input_imgs = pm_params.id_images_count;
-            auto cond_tup                   = sd_ctx->sd->cond_stage_model->get_learned_condition_with_trigger(work_ctx,
-                                                                                                               sd_ctx->sd->n_threads,
-                                                                                                               condition_params);
-            id_cond                         = std::get<0>(cond_tup);
-            class_tokens_mask               = std::get<1>(cond_tup);  //
-            struct ggml_tensor* id_embeds   = nullptr;
-            if (pmv2 && pm_params.id_embed_path != nullptr) {
-                id_embeds = load_tensor_from_file(work_ctx, pm_params.id_embed_path);
-                // print_ggml_tensor(id_embeds, true, "id_embeds:");
-            }
-            if (pmv2 && id_embeds == nullptr) {
-                LOG_WARN("Provided PhotoMaker images, but NO valid ID embeds file for PM v2");
-                LOG_WARN("Turn off PhotoMaker");
-                sd_ctx->sd->stacked_id = false;
-            } else {
-                if (pmv2 && pm_params.id_images_count != id_embeds->ne[1]) {
-                    LOG_WARN("PhotoMaker image count (%d) does NOT match ID embeds (%d). You should run face_detect.py again.", pm_params.id_images_count, id_embeds->ne[1]);
-                    LOG_WARN("Turn off PhotoMaker");
-                    sd_ctx->sd->stacked_id = false;
-                } else {
-                    id_cond.c_crossattn = sd_ctx->sd->id_encoder(work_ctx, init_img, id_cond.c_crossattn, id_embeds, class_tokens_mask);
-                    int64_t t1          = ggml_time_ms();
-                    LOG_INFO("Photomaker ID Stacking, taking %" PRId64 " ms", t1 - t0);
-                    if (sd_ctx->sd->free_params_immediately) {
-                        sd_ctx->sd->pmid_model->free_params_buffer();
-                    }
-                    // Encode input prompt without the trigger word for delayed conditioning
-                    prompt_text_only = sd_ctx->sd->cond_stage_model->remove_trigger_from_prompt(work_ctx, prompt);
-                    // printf("%s || %s \n", prompt.c_str(), prompt_text_only.c_str());
-                    prompt = prompt_text_only;  //
-                    if (sample_steps < 50) {
-                        LOG_WARN("It's recommended to use >= 50 steps for photo maker!");
-                    }
-                }
-            }
-        } else {
-            LOG_WARN("Provided PhotoMaker model file, but NO input ID images");
-            LOG_WARN("Turn off PhotoMaker");
-            sd_ctx->sd->stacked_id = false;
-        }
-    }
+    // Photo Maker
+    SDCondition id_cond = sd_ctx->sd->get_pmid_conditon(work_ctx, pm_params, condition_params);
 
     // Get learned condition
-    condition_params.text            = prompt;
     condition_params.zero_out_masked = false;
     SDCondition cond                 = sd_ctx->sd->cond_stage_model->get_learned_condition(work_ctx,
                                                                                            sd_ctx->sd->n_threads,
@@ -3131,7 +3348,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
         ggml_ext_im_set_randn_f32(noise, sd_ctx->sd->rng);
 
         int start_merge_step = -1;
-        if (sd_ctx->sd->stacked_id) {
+        if (sd_ctx->sd->use_pmid) {
             start_merge_step = int(sd_ctx->sd->pmid_model->style_strength / 100.f * sample_steps);
             // if (start_merge_step > 30)
             //     start_merge_step = 30;
@@ -3160,7 +3377,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
                                                      denoise_mask,
                                                      nullptr,
                                                      1.0f,
-                                                     easycache_params);
+                                                     cache_params);
         int64_t sampling_end    = ggml_time_ms();
         if (x_0 != nullptr) {
             // print_ggml_tensor(x_0);
@@ -3498,7 +3715,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                                                         sd_img_gen_params->increase_ref_index,
                                                         concat_latent,
                                                         denoise_mask,
-                                                        &sd_img_gen_params->easycache);
+                                                        &sd_img_gen_params->cache);
 
     size_t t2 = ggml_time_ms();
 
@@ -3869,7 +4086,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
                                  denoise_mask,
                                  vace_context,
                                  sd_vid_gen_params->vace_strength,
-                                 &sd_vid_gen_params->easycache);
+                                 &sd_vid_gen_params->cache);
 
         int64_t sampling_end = ggml_time_ms();
         LOG_INFO("sampling(high noise) completed, taking %.2fs", (sampling_end - sampling_start) * 1.0f / 1000);
@@ -3906,7 +4123,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
                                           denoise_mask,
                                           vace_context,
                                           sd_vid_gen_params->vace_strength,
-                                          &sd_vid_gen_params->easycache);
+                                          &sd_vid_gen_params->cache);
 
         int64_t sampling_end = ggml_time_ms();
         LOG_INFO("sampling completed, taking %.2fs", (sampling_end - sampling_start) * 1.0f / 1000);

stable-diffusion.h

@@ -182,6 +182,7 @@ typedef struct {
     enum prediction_t prediction;
     enum lora_apply_mode_t lora_apply_mode;
     bool offload_params_to_cpu;
+    bool enable_mmap;
     bool keep_clip_on_cpu;
     bool keep_control_net_on_cpu;
     bool keep_vae_on_cpu;
@@ -195,6 +196,7 @@ typedef struct {
     bool chroma_use_dit_mask;
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
+    bool qwen_image_zero_cond_t;
     float flow_shift;
 } sd_ctx_params_t;
 
@@ -238,12 +240,34 @@ typedef struct {
     float style_strength;
 } sd_pm_params_t;  // photo maker
 
+enum sd_cache_mode_t {
+    SD_CACHE_DISABLED = 0,
+    SD_CACHE_EASYCACHE,
+    SD_CACHE_UCACHE,
+    SD_CACHE_DBCACHE,
+    SD_CACHE_TAYLORSEER,
+    SD_CACHE_CACHE_DIT,
+};
+
 typedef struct {
-    bool enabled;
+    enum sd_cache_mode_t mode;
     float reuse_threshold;
     float start_percent;
     float end_percent;
-} sd_easycache_params_t;
+    float error_decay_rate;
+    bool use_relative_threshold;
+    bool reset_error_on_compute;
+    int Fn_compute_blocks;
+    int Bn_compute_blocks;
+    float residual_diff_threshold;
+    int max_warmup_steps;
+    int max_cached_steps;
+    int max_continuous_cached_steps;
+    int taylorseer_n_derivatives;
+    int taylorseer_skip_interval;
+    const char* scm_mask;
+    bool scm_policy_dynamic;
+} sd_cache_params_t;
 
 typedef struct {
     bool is_high_noise;
@@ -273,7 +297,7 @@ typedef struct {
     float control_strength;
     sd_pm_params_t pm_params;
     sd_tiling_params_t vae_tiling_params;
-    sd_easycache_params_t easycache;
+    sd_cache_params_t cache;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -295,7 +319,7 @@ typedef struct {
     int64_t seed;
     int video_frames;
     float vace_strength;
-    sd_easycache_params_t easycache;
+    sd_cache_params_t cache;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
@@ -325,7 +349,7 @@ SD_API enum preview_t str_to_preview(const char* str);
 SD_API const char* sd_lora_apply_mode_name(enum lora_apply_mode_t mode);
 SD_API enum lora_apply_mode_t str_to_lora_apply_mode(const char* str);
 
-SD_API void sd_easycache_params_init(sd_easycache_params_t* easycache_params);
+SD_API void sd_cache_params_init(sd_cache_params_t* cache_params);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);

ucache.hpp

@@ -0,0 +1,404 @@
+#ifndef __UCACHE_HPP__
+#define __UCACHE_HPP__
+
+#include <cmath>
+#include <limits>
+#include <unordered_map>
+#include <vector>
+
+#include "denoiser.hpp"
+#include "ggml_extend.hpp"
+
+struct UCacheConfig {
+    bool enabled                = false;
+    float reuse_threshold       = 1.0f;
+    float start_percent         = 0.15f;
+    float end_percent           = 0.95f;
+    float error_decay_rate      = 1.0f;
+    bool use_relative_threshold = true;
+    bool adaptive_threshold     = true;
+    float early_step_multiplier = 0.5f;
+    float late_step_multiplier  = 1.5f;
+    bool reset_error_on_compute = true;
+};
+
+struct UCacheCacheEntry {
+    std::vector<float> diff;
+};
+
+struct UCacheState {
+    UCacheConfig config;
+    Denoiser* denoiser                  = nullptr;
+    float start_sigma                   = std::numeric_limits<float>::max();
+    float end_sigma                     = 0.0f;
+    bool initialized                    = false;
+    bool initial_step                   = true;
+    bool skip_current_step              = false;
+    bool step_active                    = false;
+    const SDCondition* anchor_condition = nullptr;
+    std::unordered_map<const SDCondition*, UCacheCacheEntry> cache_diffs;
+    std::vector<float> prev_input;
+    std::vector<float> prev_output;
+    float output_prev_norm                = 0.0f;
+    bool has_prev_input                   = false;
+    bool has_prev_output                  = false;
+    bool has_output_prev_norm             = false;
+    bool has_relative_transformation_rate = false;
+    float relative_transformation_rate    = 0.0f;
+    float cumulative_change_rate          = 0.0f;
+    float last_input_change               = 0.0f;
+    bool has_last_input_change            = false;
+    int total_steps_skipped               = 0;
+    int current_step_index                = -1;
+    int steps_computed_since_active       = 0;
+    float accumulated_error               = 0.0f;
+    float reference_output_norm           = 0.0f;
+
+    struct BlockMetrics {
+        float sum_transformation_rate = 0.0f;
+        float sum_output_norm         = 0.0f;
+        int sample_count              = 0;
+        float min_change_rate         = std::numeric_limits<float>::max();
+        float max_change_rate         = 0.0f;
+
+        void reset() {
+            sum_transformation_rate = 0.0f;
+            sum_output_norm         = 0.0f;
+            sample_count            = 0;
+            min_change_rate         = std::numeric_limits<float>::max();
+            max_change_rate         = 0.0f;
+        }
+
+        void record(float change_rate, float output_norm) {
+            if (std::isfinite(change_rate) && change_rate > 0.0f) {
+                sum_transformation_rate += change_rate;
+                sum_output_norm += output_norm;
+                sample_count++;
+                if (change_rate < min_change_rate)
+                    min_change_rate = change_rate;
+                if (change_rate > max_change_rate)
+                    max_change_rate = change_rate;
+            }
+        }
+
+        float avg_transformation_rate() const {
+            return (sample_count > 0) ? (sum_transformation_rate / sample_count) : 0.0f;
+        }
+
+        float avg_output_norm() const {
+            return (sample_count > 0) ? (sum_output_norm / sample_count) : 0.0f;
+        }
+    };
+    BlockMetrics block_metrics;
+    int total_active_steps = 0;
+
+    void reset_runtime() {
+        initial_step      = true;
+        skip_current_step = false;
+        step_active       = false;
+        anchor_condition  = nullptr;
+        cache_diffs.clear();
+        prev_input.clear();
+        prev_output.clear();
+        output_prev_norm                 = 0.0f;
+        has_prev_input                   = false;
+        has_prev_output                  = false;
+        has_output_prev_norm             = false;
+        has_relative_transformation_rate = false;
+        relative_transformation_rate     = 0.0f;
+        cumulative_change_rate           = 0.0f;
+        last_input_change                = 0.0f;
+        has_last_input_change            = false;
+        total_steps_skipped              = 0;
+        current_step_index               = -1;
+        steps_computed_since_active      = 0;
+        accumulated_error                = 0.0f;
+        reference_output_norm            = 0.0f;
+        block_metrics.reset();
+        total_active_steps = 0;
+    }
+
+    void init(const UCacheConfig& cfg, Denoiser* d) {
+        config      = cfg;
+        denoiser    = d;
+        initialized = cfg.enabled && d != nullptr;
+        reset_runtime();
+        if (initialized) {
+            start_sigma = percent_to_sigma(config.start_percent);
+            end_sigma   = percent_to_sigma(config.end_percent);
+        }
+    }
+
+    void set_sigmas(const std::vector<float>& sigmas) {
+        if (!initialized || sigmas.size() < 2) {
+            return;
+        }
+        size_t n_steps = sigmas.size() - 1;
+
+        size_t start_step = static_cast<size_t>(config.start_percent * n_steps);
+        size_t end_step   = static_cast<size_t>(config.end_percent * n_steps);
+
+        if (start_step >= n_steps)
+            start_step = n_steps - 1;
+        if (end_step >= n_steps)
+            end_step = n_steps - 1;
+
+        start_sigma = sigmas[start_step];
+        end_sigma   = sigmas[end_step];
+
+        if (start_sigma < end_sigma) {
+            std::swap(start_sigma, end_sigma);
+        }
+    }
+
+    bool enabled() const {
+        return initialized && config.enabled;
+    }
+
+    float percent_to_sigma(float percent) const {
+        if (!denoiser) {
+            return 0.0f;
+        }
+        if (percent <= 0.0f) {
+            return std::numeric_limits<float>::max();
+        }
+        if (percent >= 1.0f) {
+            return 0.0f;
+        }
+        float t = (1.0f - percent) * (TIMESTEPS - 1);
+        return denoiser->t_to_sigma(t);
+    }
+
+    void begin_step(int step_index, float sigma) {
+        if (!enabled()) {
+            return;
+        }
+        if (step_index == current_step_index) {
+            return;
+        }
+        current_step_index    = step_index;
+        skip_current_step     = false;
+        has_last_input_change = false;
+        step_active           = false;
+
+        if (sigma > start_sigma) {
+            return;
+        }
+        if (!(sigma > end_sigma)) {
+            return;
+        }
+        step_active = true;
+        total_active_steps++;
+    }
+
+    bool step_is_active() const {
+        return enabled() && step_active;
+    }
+
+    bool is_step_skipped() const {
+        return enabled() && step_active && skip_current_step;
+    }
+
+    float get_adaptive_threshold(int estimated_total_steps = 0) const {
+        float base_threshold = config.reuse_threshold;
+
+        if (!config.adaptive_threshold) {
+            return base_threshold;
+        }
+
+        int effective_total = estimated_total_steps;
+        if (effective_total <= 0) {
+            effective_total = std::max(20, steps_computed_since_active * 2);
+        }
+
+        float progress = (effective_total > 0) ? (static_cast<float>(steps_computed_since_active) / effective_total) : 0.0f;
+
+        float multiplier = 1.0f;
+        if (progress < 0.2f) {
+            multiplier = config.early_step_multiplier;
+        } else if (progress > 0.8f) {
+            multiplier = config.late_step_multiplier;
+        }
+
+        return base_threshold * multiplier;
+    }
+
+    bool has_cache(const SDCondition* cond) const {
+        auto it = cache_diffs.find(cond);
+        return it != cache_diffs.end() && !it->second.diff.empty();
+    }
+
+    void update_cache(const SDCondition* cond, ggml_tensor* input, ggml_tensor* output) {
+        UCacheCacheEntry& entry = cache_diffs[cond];
+        size_t ne               = static_cast<size_t>(ggml_nelements(output));
+        entry.diff.resize(ne);
+        float* out_data = (float*)output->data;
+        float* in_data  = (float*)input->data;
+
+        for (size_t i = 0; i < ne; ++i) {
+            entry.diff[i] = out_data[i] - in_data[i];
+        }
+    }
+
+    void apply_cache(const SDCondition* cond, ggml_tensor* input, ggml_tensor* output) {
+        auto it = cache_diffs.find(cond);
+        if (it == cache_diffs.end() || it->second.diff.empty()) {
+            return;
+        }
+
+        copy_ggml_tensor(output, input);
+        float* out_data                = (float*)output->data;
+        const std::vector<float>& diff = it->second.diff;
+        for (size_t i = 0; i < diff.size(); ++i) {
+            out_data[i] += diff[i];
+        }
+    }
+
+    bool before_condition(const SDCondition* cond,
+                          ggml_tensor* input,
+                          ggml_tensor* output,
+                          float sigma,
+                          int step_index) {
+        if (!enabled() || step_index < 0) {
+            return false;
+        }
+        if (step_index != current_step_index) {
+            begin_step(step_index, sigma);
+        }
+        if (!step_active) {
+            return false;
+        }
+
+        if (initial_step) {
+            anchor_condition = cond;
+            initial_step     = false;
+        }
+
+        bool is_anchor = (cond == anchor_condition);
+
+        if (skip_current_step) {
+            if (has_cache(cond)) {
+                apply_cache(cond, input, output);
+                return true;
+            }
+            return false;
+        }
+
+        if (!is_anchor) {
+            return false;
+        }
+
+        if (!has_prev_input || !has_prev_output || !has_cache(cond)) {
+            return false;
+        }
+
+        size_t ne = static_cast<size_t>(ggml_nelements(input));
+        if (prev_input.size() != ne) {
+            return false;
+        }
+
+        float* input_data = (float*)input->data;
+        last_input_change = 0.0f;
+        for (size_t i = 0; i < ne; ++i) {
+            last_input_change += std::fabs(input_data[i] - prev_input[i]);
+        }
+        if (ne > 0) {
+            last_input_change /= static_cast<float>(ne);
+        }
+        has_last_input_change = true;
+
+        if (has_output_prev_norm && has_relative_transformation_rate &&
+            last_input_change > 0.0f && output_prev_norm > 0.0f) {
+            float approx_output_change_rate = (relative_transformation_rate * last_input_change) / output_prev_norm;
+            accumulated_error               = accumulated_error * config.error_decay_rate + approx_output_change_rate;
+
+            float effective_threshold = get_adaptive_threshold();
+            if (config.use_relative_threshold && reference_output_norm > 0.0f) {
+                effective_threshold = effective_threshold * reference_output_norm;
+            }
+
+            if (accumulated_error < effective_threshold) {
+                skip_current_step = true;
+                total_steps_skipped++;
+                apply_cache(cond, input, output);
+                return true;
+            } else if (config.reset_error_on_compute) {
+                accumulated_error = 0.0f;
+            }
+        }
+
+        return false;
+    }
+
+    void after_condition(const SDCondition* cond, ggml_tensor* input, ggml_tensor* output) {
+        if (!step_is_active()) {
+            return;
+        }
+
+        update_cache(cond, input, output);
+
+        if (cond != anchor_condition) {
+            return;
+        }
+
+        size_t ne      = static_cast<size_t>(ggml_nelements(input));
+        float* in_data = (float*)input->data;
+        prev_input.resize(ne);
+        for (size_t i = 0; i < ne; ++i) {
+            prev_input[i] = in_data[i];
+        }
+        has_prev_input = true;
+
+        float* out_data     = (float*)output->data;
+        float output_change = 0.0f;
+        if (has_prev_output && prev_output.size() == ne) {
+            for (size_t i = 0; i < ne; ++i) {
+                output_change += std::fabs(out_data[i] - prev_output[i]);
+            }
+            if (ne > 0) {
+                output_change /= static_cast<float>(ne);
+            }
+        }
+
+        prev_output.resize(ne);
+        for (size_t i = 0; i < ne; ++i) {
+            prev_output[i] = out_data[i];
+        }
+        has_prev_output = true;
+
+        float mean_abs = 0.0f;
+        for (size_t i = 0; i < ne; ++i) {
+            mean_abs += std::fabs(out_data[i]);
+        }
+        output_prev_norm     = (ne > 0) ? (mean_abs / static_cast<float>(ne)) : 0.0f;
+        has_output_prev_norm = output_prev_norm > 0.0f;
+
+        if (reference_output_norm == 0.0f) {
+            reference_output_norm = output_prev_norm;
+        }
+
+        if (has_last_input_change && last_input_change > 0.0f && output_change > 0.0f) {
+            float rate = output_change / last_input_change;
+            if (std::isfinite(rate)) {
+                relative_transformation_rate     = rate;
+                has_relative_transformation_rate = true;
+                block_metrics.record(rate, output_prev_norm);
+            }
+        }
+
+        has_last_input_change = false;
+    }
+
+    void log_block_metrics() const {
+        if (block_metrics.sample_count > 0) {
+            LOG_INFO("UCacheBlockMetrics: samples=%d, avg_rate=%.4f, min=%.4f, max=%.4f, avg_norm=%.4f",
+                     block_metrics.sample_count,
+                     block_metrics.avg_transformation_rate(),
+                     block_metrics.min_change_rate,
+                     block_metrics.max_change_rate,
+                     block_metrics.avg_output_norm());
+        }
+    }
+};
+
+#endif  // __UCACHE_HPP__

util.cpp

@@ -95,9 +95,71 @@ bool is_directory(const std::string& path) {
     return (attributes != INVALID_FILE_ATTRIBUTES && (attributes & FILE_ATTRIBUTE_DIRECTORY));
 }
 
+class MmapWrapperImpl : public MmapWrapper {
+public:
+    MmapWrapperImpl(void* data, size_t size, HANDLE hfile, HANDLE hmapping)
+        : MmapWrapper(data, size), hfile_(hfile), hmapping_(hmapping) {}
+
+    ~MmapWrapperImpl() override {
+        UnmapViewOfFile(data_);
+        CloseHandle(hmapping_);
+        CloseHandle(hfile_);
+    }
+
+private:
+    HANDLE hfile_;
+    HANDLE hmapping_;
+};
+
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
+    void* mapped_data = nullptr;
+    size_t file_size  = 0;
+
+    HANDLE file_handle = CreateFileA(
+        filename.c_str(),
+        GENERIC_READ,
+        FILE_SHARE_READ,
+        NULL,
+        OPEN_EXISTING,
+        FILE_ATTRIBUTE_NORMAL,
+        NULL);
+
+    if (file_handle == INVALID_HANDLE_VALUE) {
+        return nullptr;
+    }
+
+    LARGE_INTEGER size;
+    if (!GetFileSizeEx(file_handle, &size)) {
+        CloseHandle(file_handle);
+        return nullptr;
+    }
+
+    file_size = static_cast<size_t>(size.QuadPart);
+
+    HANDLE mapping_handle = CreateFileMapping(file_handle, NULL, PAGE_READONLY, 0, 0, NULL);
+
+    if (mapping_handle == NULL) {
+        CloseHandle(file_handle);
+        return nullptr;
+    }
+
+    mapped_data = MapViewOfFile(mapping_handle, FILE_MAP_READ, 0, 0, file_size);
+
+    if (mapped_data == NULL) {
+        CloseHandle(mapping_handle);
+        CloseHandle(file_handle);
+        return nullptr;
+    }
+
+    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size, file_handle, mapping_handle);
+}
+
 #else  // Unix
 #include <dirent.h>
+#include <fcntl.h>
+#include <sys/mman.h>
 #include <sys/stat.h>
+#include <unistd.h>
 
 bool file_exists(const std::string& filename) {
     struct stat buffer;
@@ -109,8 +171,64 @@ bool is_directory(const std::string& path) {
     return (stat(path.c_str(), &buffer) == 0 && S_ISDIR(buffer.st_mode));
 }
 
+class MmapWrapperImpl : public MmapWrapper {
+public:
+    MmapWrapperImpl(void* data, size_t size)
+        : MmapWrapper(data, size) {}
+
+    ~MmapWrapperImpl() override {
+        munmap(data_, size_);
+    }
+};
+
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
+    int file_descriptor = open(filename.c_str(), O_RDONLY);
+    if (file_descriptor == -1) {
+        return nullptr;
+    }
+
+    int mmap_flags = MAP_PRIVATE;
+
+#ifdef __linux__
+    // performance flags used by llama.cpp
+    // posix_fadvise(file_descriptor, 0, 0, POSIX_FADV_SEQUENTIAL);
+    // mmap_flags |= MAP_POPULATE;
 #endif
 
+    struct stat sb;
+    if (fstat(file_descriptor, &sb) == -1) {
+        close(file_descriptor);
+        return nullptr;
+    }
+
+    size_t file_size = sb.st_size;
+
+    void* mapped_data = mmap(NULL, file_size, PROT_READ, mmap_flags, file_descriptor, 0);
+
+    close(file_descriptor);
+
+    if (mapped_data == MAP_FAILED) {
+        return nullptr;
+    }
+
+#ifdef __linux__
+    // performance flags used by llama.cpp
+    // posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
+#endif
+
+    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size);
+}
+
+#endif
+
+bool MmapWrapper::copy_data(void* buf, size_t n, size_t offset) const {
+    if (offset >= size_ || n > (size_ - offset)) {
+        return false;
+    }
+    std::memcpy(buf, data() + offset, n);
+    return true;
+}
+
 // get_num_physical_cores is copy from
 // https://github.com/ggerganov/llama.cpp/blob/master/examples/common.cpp
 // LICENSE: https://github.com/ggerganov/llama.cpp/blob/master/LICENSE

util.h

@@ -2,6 +2,7 @@
 #define __UTIL_H__
 
 #include <cstdint>
+#include <memory>
 #include <string>
 #include <vector>
 
@@ -43,6 +44,28 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
 
 sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height);
 
+class MmapWrapper {
+public:
+    static std::unique_ptr<MmapWrapper> create(const std::string& filename);
+
+    virtual ~MmapWrapper() = default;
+
+    MmapWrapper(const MmapWrapper&)            = delete;
+    MmapWrapper& operator=(const MmapWrapper&) = delete;
+    MmapWrapper(MmapWrapper&&)                 = delete;
+    MmapWrapper& operator=(MmapWrapper&&)      = delete;
+
+    const uint8_t* data() const { return static_cast<uint8_t*>(data_); }
+    size_t size() const { return size_; }
+    bool copy_data(void* buf, size_t n, size_t offset) const;
+
+protected:
+    MmapWrapper(void* data, size_t size)
+        : data_(data), size_(size) {}
+    void* data_  = nullptr;
+    size_t size_ = 0;
+};
+
 std::string path_join(const std::string& p1, const std::string& p2);
 std::vector<std::string> split_string(const std::string& str, char delimiter);
 void pretty_progress(int step, int steps, float time);