merge branch 'master' into ltx2.3

.github/pull_request_template.md

@@ -0,0 +1,15 @@
+## Summary
+
+<!-- Describe what changed and why. Keep the PR focused on one clear change. -->
+
+## Related Issue / Discussion
+
+<!-- Link related issues, discussions, or previous PRs if applicable. -->
+
+## Additional Information
+
+<!-- Add verification notes, screenshots, sample output, or other context when applicable. -->
+
+## Checklist
+
+- [ ] I have read and confirmed this PR follows the [contribution guidelines](https://github.com/leejet/stable-diffusion.cpp/blob/master/CONTRIBUTING.md).

CMakeLists.txt

@@ -13,7 +13,9 @@ if (MSVC)
     add_compile_definitions(_SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING)
     add_compile_options(
         $<$<COMPILE_LANGUAGE:C>:/MP>
+        $<$<COMPILE_LANGUAGE:C>:/utf-8>
         $<$<COMPILE_LANGUAGE:CXX>:/MP>
+        $<$<COMPILE_LANGUAGE:CXX>:/utf-8>
     )
 endif()
 
@@ -69,6 +71,12 @@ option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" O
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
 if(SD_CUDA)
     message("-- Use CUDA as backend stable-diffusion")
     set(GGML_CUDA ON)
@@ -106,19 +114,28 @@ if(SD_WEBP)
           "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
     endif()
     if(SD_USE_SYSTEM_WEBP)
-        find_package(WebP REQUIRED)
-        add_library(webp ALIAS WebP::webp)
-        # libwebp CMake target naming is not consistent across versions/distros.
-        # Some export WebP::libwebpmux, others export WebP::webpmux.
-        if(TARGET WebP::libwebpmux)
-            add_library(libwebpmux ALIAS WebP::libwebpmux)
-        elseif(TARGET WebP::webpmux)
-            add_library(libwebpmux ALIAS WebP::webpmux)
+        find_package(WebP)
+        if(WebP_FOUND)
+            add_library(webp ALIAS WebP::webp)
+            # libwebp CMake target naming is not consistent across versions/distros.
+            # Some export WebP::libwebpmux, others export WebP::webpmux.
+            if(TARGET WebP::libwebpmux)
+                add_library(libwebpmux ALIAS WebP::libwebpmux)
+            elseif(TARGET WebP::webpmux)
+                add_library(libwebpmux ALIAS WebP::webpmux)
+            else()
+                message(FATAL_ERROR
+                    "Could not find a compatible webpmux target in system WebP package. "
+                    "Expected WebP::libwebpmux or WebP::webpmux."
+                )
+            endif()
         else()
-            message(FATAL_ERROR
-                "Could not find a compatible webpmux target in system WebP package. "
-                "Expected WebP::libwebpmux or WebP::webpmux."
-            )
+            find_package(PkgConfig REQUIRED)
+            pkg_check_modules(WebP REQUIRED IMPORTED_TARGET GLOBAL libwebp)
+            pkg_check_modules(WebPMux REQUIRED IMPORTED_TARGET GLOBAL libwebpmux)
+            link_libraries(PkgConfig::WebP)
+            link_libraries(PkgConfig::WebPMux)
+            add_library(libwebpmux ALIAS PkgConfig::WebPMux)
         endif()
     endif()
 endif()
@@ -133,18 +150,26 @@ if(SD_WEBM)
           "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
     endif()
     if(SD_USE_SYSTEM_WEBM)
-        find_path(WEBM_INCLUDE_DIR
-            NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
-            PATH_SUFFIXES webm
-            REQUIRED)
-        find_library(WEBM_LIBRARY
-            NAMES webm libwebm
-            REQUIRED)
+        find_package(PkgConfig)
+        if(PkgConfig_FOUND)
+            pkg_check_modules(WebM REQUIRED IMPORTED_TARGET GLOBAL libwebm)
+        endif()
+        if(PkgConfig_FOUND AND WebM_FOUND)
+            link_libraries(PkgConfig::WebM)
+        else()
+            find_path(WEBM_INCLUDE_DIR
+                NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
+                PATH_SUFFIXES webm
+                REQUIRED)
+            find_library(WEBM_LIBRARY
+                NAMES webm libwebm
+                REQUIRED)
 
-        add_library(webm UNKNOWN IMPORTED)
-        set_target_properties(webm PROPERTIES
-            IMPORTED_LOCATION "${WEBM_LIBRARY}"
-            INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+            add_library(webm UNKNOWN IMPORTED)
+            set_target_properties(webm PROPERTIES
+                IMPORTED_LOCATION "${WEBM_LIBRARY}"
+                INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+        endif()
     endif()
 endif()
 

CONTRIBUTING.md

@@ -0,0 +1,65 @@
+# Contributing
+
+This document collects general contribution conventions for this repository.
+
+## Before You Start
+
+Before opening a PR, please search existing PRs to avoid duplicating ongoing work.
+
+For large-scale refactors or changes with broad impact, please open an issue first to discuss the approach before submitting a PR.
+
+If you want to update a third-party dependency, please open an issue first instead of submitting a direct PR. See [Dependency Updates](#dependency-updates) for details.
+
+## Pull Requests
+
+Keep each PR focused on one clear change. Large or overly complex PRs are harder to review and may not be merged.
+
+Follow Conventional Commit-style subjects seen in history: `feat:`, `fix:`, `refactor:`, `ci:`, `docs:`, `chore:`. Keep subjects imperative and scoped.
+
+PRs should include:
+
+- What changed and why (short problem/solution summary).
+- Verification evidence when applicable (commands and key outputs).
+- Linked issue/PR context when applicable.
+- Screenshots or sample outputs for UI/visual behavior changes.
+
+## Code Style
+
+Format code according to the repository style before submitting changes.
+
+Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `format-code.sh` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
+
+Naming conventions:
+
+- Use `PascalCase` for class/struct/type names.
+- In `PascalCase` names, preserve common abbreviations in uppercase, for example `SD`, `API`, `HTTP`, `JSON`, `RGB`, `VAE`, `TAE`, `LoRA`, and `WebP`.
+- Use `snake_case` for functions, methods, variables, and file names unless an existing API requires a different style.
+- Use a trailing underscore for private data member names, for example `hidden_size_` or `tokenizer_`.
+- Use `.h` for C and C++ header files. Do not introduce new `.hpp` headers.
+- Use macro-based header include guards instead of `#pragma once`.
+- Format header include guards as `__SD_{PATH}__`, where `{PATH}` is the header path in uppercase snake case without the file extension. For example, `src/sample.h` should use `__SD_SAMPLE_H__`.
+- Do not introduce anonymous namespaces in new or modified code; prefer `static` file-local functions/variables or an explicit named namespace when scoping is needed.
+- In `class`/`struct` definitions, place data members before member functions unless an existing type already clearly follows a different pattern.
+- Keep `test_*.cpp` / `test_*.py` naming for tests.
+
+Some older code in the project may not fully follow the current conventions. Please do not submit PRs that only rewrite existing code to match style rules.
+
+## AI-Assisted Contributions
+
+AI tools may be used to assist development, but contributors are responsible for the quality and correctness of the submitted code.
+
+If any part of a contribution was generated with AI assistance, the contributor must perform a thorough human review before submitting the PR and understand every changed line.
+
+Do not list AI tools as co-authors. The human contributor is the sole responsible author of the submitted code.
+
+Please do not submit AI-generated code that you do not understand, and do not include meaningless experiments, temporary test code, or unrelated generated output in a PR.
+
+## Dependency Updates
+
+Do not submit PRs that update `ggml`. `ggml` updates are performed only after local validation by the maintainer.
+
+Other third-party dependencies are not updated unless necessary. If you want to update a dependency, please open an issue first instead of submitting a direct PR.
+
+## Security & Configuration
+
+Do not commit model weights, secrets, or local absolute paths. Keep large binaries out of git unless intentionally tracked release assets.

README.md

@@ -58,6 +58,7 @@ API and command-line option may change frequently.***
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
+    - [HiDream-O1-Image](./docs/hidream_o1_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -148,6 +149,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [Ovis-Image](./docs/ovis_image.md)
 - [Anima](./docs/anima.md)
 - [ERNIE-Image](./docs/ernie_image.md)
+- [HiDream-O1-Image](./docs/hidream_o1_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -163,6 +165,7 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 
 * Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
+* Golang (non-cgo): [l8bloom/gosd](https://github.com/l8bloom/gosd)
 * C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
 * Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
 * Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)

docs/hidream_o1_image.md

@@ -0,0 +1,20 @@
+# How to Use
+
+## Download weights
+
+- Download HiDream-O1-Image-Dev
+    - safetensors: https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints
+- Download HiDream-O1-Image
+    - safetensors: https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints
+
+## Examples
+
+### HiDream-O1-Image-Dev
+
+```
+.\bin\Release\sd-cli.exe -m  ..\..\ComfyUI\models\diffusion_models\hidream_o1_image_dev_bf16.safetensors -p "a lovely cat holding a sign says 
+'hidream o1 cpp'" --cfg-scale 1.0  -v -H 1024 -W 1024
+```
+
+<img width="256" alt="HiDream-O1-Image-Dev example" src="../assets/hidream-o1/dev_example.png" />
+

examples/cli/README.md

@@ -55,7 +55,7 @@ Context Options:
                                            then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
   --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
-                                           graph splitting
+                                           graph splitting; -1 auto-detects free VRAM minus 1 GiB
   --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
@@ -103,6 +103,8 @@ Generation Options:
   --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                            (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                            antialiased), or a model name under --hires-upscalers-dir (default: Latent)
+  --extra-sample-args <string>             extra sampler args, key=value list. Currently lcm supports noise_clip_std,
+                                           noise_scale_start, noise_scale_end
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
@@ -163,10 +165,10 @@ Generation Options:
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
                                            dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
-                                           er_sde] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
   --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
                                            dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
-                                           res_2s, er_sde] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] 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, bong_tangent], default:
                                            discrete

examples/common/common.cpp

@@ -405,7 +405,7 @@ ArgOptions SDContextParams::get_options() {
     options.float_options = {
         {"",
          "--max-vram",
-         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting",
+         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; -1 auto-detects free VRAM minus 1 GiB",
          &max_vram},
     };
 
@@ -819,6 +819,10 @@ ArgOptions SDGenerationParams::get_options() {
          "Latent (antialiased), Latent (bicubic), Latent (bicubic antialiased), or a model name "
          "under --hires-upscalers-dir (default: Latent)",
          &hires_upscaler},
+        {"",
+         "--extra-sample-args",
+         "extra sampler args, key=value list. Currently lcm supports noise_clip_std, noise_scale_start, noise_scale_end",
+         &extra_sample_args},
     };
 
     options.int_options = {
@@ -1261,12 +1265,12 @@ ArgOptions SDGenerationParams::get_options() {
          on_seed_arg},
         {"",
          "--sampling-method",
-         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde] "
+         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
          on_sample_method_arg},
         {"",
          "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde]"
+         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          " default: euler for Flux/SD3/Wan, euler_a otherwise",
          on_high_noise_sample_method_arg},
         {"",
@@ -1624,6 +1628,7 @@ bool SDGenerationParams::from_json_str(
 
     auto parse_sample_params_json = [&](const json& sample_json,
                                         sd_sample_params_t& target_params,
+                                        std::string& target_extra_sample_args,
                                         std::vector<int>& target_skip_layers,
                                         std::vector<float>* target_custom_sigmas) {
         if (sample_json.contains("sample_steps") && sample_json["sample_steps"].is_number_integer()) {
@@ -1638,6 +1643,9 @@ bool SDGenerationParams::from_json_str(
         if (sample_json.contains("flow_shift") && sample_json["flow_shift"].is_number()) {
             target_params.flow_shift = sample_json["flow_shift"];
         }
+        if (sample_json.contains("extra_sample_args") && sample_json["extra_sample_args"].is_string()) {
+            target_extra_sample_args = sample_json["extra_sample_args"].get<std::string>();
+        }
         if (target_custom_sigmas != nullptr &&
             sample_json.contains("custom_sigmas") &&
             sample_json["custom_sigmas"].is_array()) {
@@ -1685,11 +1693,12 @@ bool SDGenerationParams::from_json_str(
     };
 
     if (j.contains("sample_params") && j["sample_params"].is_object()) {
-        parse_sample_params_json(j["sample_params"], sample_params, skip_layers, &custom_sigmas);
+        parse_sample_params_json(j["sample_params"], sample_params, extra_sample_args, skip_layers, &custom_sigmas);
     }
     if (j.contains("high_noise_sample_params") && j["high_noise_sample_params"].is_object()) {
         parse_sample_params_json(j["high_noise_sample_params"],
                                  high_noise_sample_params,
+                                 high_noise_extra_sample_args,
                                  high_noise_skip_layers,
                                  nullptr);
     }
@@ -2119,6 +2128,8 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
     sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
+    sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
+    high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     sd_pm_params_t pm_params = {
@@ -2188,6 +2199,8 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
     sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
+    sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
+    high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     params.loras                    = lora_vec.empty() ? nullptr : lora_vec.data();
@@ -2328,6 +2341,7 @@ static json build_sampling_metadata_json(const sd_sample_params_t& sample_params
         {"eta", sample_params.eta},
         {"shifted_timestep", sample_params.shifted_timestep},
         {"flow_shift", sample_params.flow_shift},
+        {"extra_sample_args", safe_json_string(sample_params.extra_sample_args)},
         {"guidance",
          {
              {"txt_cfg", sample_params.guidance.txt_cfg},
@@ -2519,6 +2533,9 @@ std::string get_image_params(const SDContextParams& ctx_params,
     }
     parameter_string += "Guidance: " + std::to_string(gen_params.sample_params.guidance.distilled_guidance) + ", ";
     parameter_string += "Eta: " + std::to_string(gen_params.sample_params.eta) + ", ";
+    if (!gen_params.extra_sample_args.empty()) {
+        parameter_string += "Extra sample args: " + gen_params.extra_sample_args + ", ";
+    }
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(gen_params.get_resolved_width()) + "x" + std::to_string(gen_params.get_resolved_height()) + ", ";
     parameter_string += "Model: " + sd_basename(ctx_params.model_path) + ", ";

examples/common/common.h

@@ -170,6 +170,8 @@ struct SDGenerationParams {
 
     sd_sample_params_t sample_params;
     sd_sample_params_t high_noise_sample_params;
+    std::string extra_sample_args;
+    std::string high_noise_extra_sample_args;
     std::vector<int> skip_layers            = {7, 8, 9};
     std::vector<int> high_noise_skip_layers = {7, 8, 9};
 

examples/server/README.md

@@ -157,7 +157,7 @@ Context Options:
                                            then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
   --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
-                                           graph splitting
+                                           graph splitting; -1 auto-detects free VRAM minus 1 GiB
   --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
@@ -205,6 +205,8 @@ Default Generation Options:
   --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                            (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                            antialiased), or a model name under --hires-upscalers-dir (default: Latent)
+  --extra-sample-args <string>             extra sampler args, key=value list. Currently lcm supports noise_clip_std,
+                                           noise_scale_start, noise_scale_end
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
@@ -264,10 +266,10 @@ Default Generation Options:
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
                                            dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
-                                           er_sde] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
   --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
                                            dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
-                                           res_2s, er_sde] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] 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, bong_tangent], default:
                                            discrete

examples/server/main.cpp

@@ -145,7 +145,7 @@ int main(int argc, const char** argv) {
     register_sdapi_endpoints(svr, runtime);
     register_sdcpp_api_endpoints(svr, runtime);
 
-    LOG_INFO("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
+    LOG_INFO("listening on: http://%s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
     svr.listen(svr_params.listen_ip, svr_params.listen_port);
 
     {

examples/server/routes_sdapi.cpp

@@ -67,6 +67,10 @@ static enum sample_method_t get_sdapi_sample_method(std::string name) {
         {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
         {"res 2s", RES_2S_SAMPLE_METHOD},
         {"k_res_2s", RES_2S_SAMPLE_METHOD},
+        {"euler_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"k_euler_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"euler_a_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"k_euler_a_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
     };
     auto it = hardcoded.find(name);
     return it != hardcoded.end() ? it->second : SAMPLE_METHOD_COUNT;

include/stable-diffusion.h

@@ -51,6 +51,8 @@ enum sample_method_t {
     RES_MULTISTEP_SAMPLE_METHOD,
     RES_2S_SAMPLE_METHOD,
     ER_SDE_SAMPLE_METHOD,
+    EULER_CFG_PP_SAMPLE_METHOD,
+    EULER_A_CFG_PP_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -206,7 +208,7 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
-    float max_vram;
+    float max_vram;  // GiB budget for graph-cut segmented param offload (0 = disabled, -1 = auto free VRAM minus 1 GiB)
 } sd_ctx_params_t;
 
 typedef struct {
@@ -248,6 +250,7 @@ typedef struct {
     float* custom_sigmas;
     int custom_sigmas_count;
     float flow_shift;
+    const char* extra_sample_args;
 } sd_sample_params_t;
 
 typedef struct {

src/conditioner.hpp

@@ -16,6 +16,12 @@ struct SDCondition {
     sd::Tensor<float> c_concat;
     sd::Tensor<int32_t> c_t5_ids;
     sd::Tensor<float> c_t5_weights;
+    sd::Tensor<int32_t> c_input_ids;
+    sd::Tensor<int32_t> c_position_ids;
+    sd::Tensor<int32_t> c_token_types;
+    sd::Tensor<int32_t> c_vinput_mask;
+    std::vector<std::pair<int, sd::Tensor<float>>> c_image_embeds;
+    std::vector<sd::Tensor<float>> c_ref_images;
 
     std::vector<sd::Tensor<float>> extra_c_crossattns;
 
@@ -28,10 +34,24 @@ struct SDCondition {
 
     bool empty() const {
         if (!c_crossattn.empty() || !c_vector.empty() || !c_concat.empty() ||
-            !c_t5_ids.empty() || !c_t5_weights.empty()) {
+            !c_t5_ids.empty() || !c_t5_weights.empty() ||
+            !c_input_ids.empty() || !c_position_ids.empty() ||
+            !c_token_types.empty() || !c_vinput_mask.empty()) {
             return false;
         }
 
+        for (const auto& image_embed : c_image_embeds) {
+            if (!image_embed.second.empty()) {
+                return false;
+            }
+        }
+
+        for (const auto& tensor : c_ref_images) {
+            if (!tensor.empty()) {
+                return false;
+            }
+        }
+
         for (const auto& tensor : extra_c_crossattns) {
             if (!tensor.empty()) {
                 return false;

src/denoiser.hpp

@@ -2,6 +2,7 @@
 #define __DENOISER_HPP__
 
 #include <cmath>
+#include <string>
 #include <utility>
 
 #include "ggml_extend.hpp"
@@ -752,7 +753,7 @@ struct Flux2FlowDenoiser : public FluxFlowDenoiser {
     }
 };
 
-typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int)> denoise_cb_t;
+typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int, sd::Tensor<float>*)> denoise_cb_t;
 
 static std::pair<float, float> get_ancestral_step(float sigma_from,
                                                   float sigma_to,
@@ -823,45 +824,33 @@ static std::tuple<float, float, float> get_ancestral_step(float sigma_from,
 static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
                                                 sd::Tensor<float> x,
                                                 const std::vector<float>& sigmas,
-                                                std::shared_ptr<RNG> rng,
-                                                float eta) {
+                                                std::shared_ptr<RNG> rng = nullptr,
+                                                bool is_flow_denoiser    = false,
+                                                float eta                = 0.f) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
+        float sigma_to    = sigmas[i + 1];
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
-        sd::Tensor<float> denoised  = std::move(denoised_opt);
-        sd::Tensor<float> d         = (x - denoised) / sigma;
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
-        x += d * (sigma_down - sigmas[i]);
-        if (sigmas[i + 1] > 0) {
-            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
-        }
-    }
-    return x;
-}
-
-static sd::Tensor<float> sample_euler_flow(denoise_cb_t model,
-                                           sd::Tensor<float> x,
-                                           const std::vector<float>& sigmas,
-                                           std::shared_ptr<RNG> rng,
-                                           float eta) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
-            return {};
-        }
-        sd::Tensor<float> denoised               = std::move(denoised_opt);
-        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step_flow(sigma, sigmas[i + 1], eta);
-        float sigma_ratio                        = sigma_down / sigma;
-        x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
-
-        if (sigma_up > 0.0f) {
-            x = alpha_scale * x + sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        if (sigma_to == 0.f) {
+            x = denoised;
+        } else if (eta == 0.f) {
+            float sigma_ratio = sigma_to / sigma;
+            x                 = sigma_ratio * x + (1.0 - sigma_ratio) * denoised;
+        } else {
+            auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma, sigma_to, eta, is_flow_denoiser);
+            float sigma_ratio                        = sigma_down / sigma;
+            x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
+            if (sigma_up > 0.f) {
+                if (is_flow_denoiser) {
+                    x *= alpha_scale;
+                }
+                x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+            }
         }
     }
     return x;
@@ -873,7 +862,7 @@ static sd::Tensor<float> sample_euler(denoise_cb_t model,
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -889,7 +878,7 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
                                      const std::vector<float>& sigmas) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1));
+        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -900,7 +889,7 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
             x += d * dt;
         } else {
             sd::Tensor<float> x2 = x + d * dt;
-            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1);
+            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -917,7 +906,7 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
                                      const std::vector<float>& sigmas) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1));
+        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -930,7 +919,7 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
             float dt_1           = sigma_mid - sigmas[i];
             float dt_2           = sigmas[i + 1] - sigmas[i];
             sd::Tensor<float> x2 = x + d * dt_1;
-            auto denoised2_opt   = model(x2, sigma_mid, i + 1);
+            auto denoised2_opt   = model(x2, sigma_mid, i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -951,7 +940,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1));
+        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -967,7 +956,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
             float s              = t + 0.5f * h;
             float sigma_s        = sigma_fn(s);
             sd::Tensor<float> x2 = (sigma_s / sigma_fn(t)) * x - (exp(-h * 0.5f) - 1) * denoised;
-            auto denoised2_opt   = model(x2, sigma_s, i + 1);
+            auto denoised2_opt   = model(x2, sigma_s, i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -994,7 +983,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
 
         bool opt_first_step = (1.0 - sigma < 1e-6);
 
-        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1));
+        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1023,8 +1012,8 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                 // so sigma_s = 1 = sigma, and sigma_s_i_ratio = sigma_s / sigma = 1
                 // u = (x*sigma_s_i_ratio)+(denoised*(1.0f-sigma_s_i_ratio))
                 //   = (x*1)+(denoised*0) = x
-                // so D_i = model(u, sigma_s, i + 1)
-                //        = model(x, sigma,   i + 1)
+                // so D_i = model(u, sigma_s, i + 1, nullptr)
+                //        = model(x, sigma,   i + 1, nullptr)
                 //        = denoised
                 D_i = denoised;
 
@@ -1057,7 +1046,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                 float sigma_s_i_ratio = sigma_s / sigma;
                 sd::Tensor<float> u   = (x * sigma_s_i_ratio) + (denoised * (1.0f - sigma_s_i_ratio));
 
-                auto denoised2_opt = model(u, sigma_s, i + 1);
+                auto denoised2_opt = model(u, sigma_s, i + 1, nullptr);
                 if (denoised2_opt.empty()) {
                     return {};
                 }
@@ -1084,7 +1073,7 @@ static sd::Tensor<float> sample_dpmpp_2m(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1116,7 +1105,7 @@ static sd::Tensor<float> sample_dpmpp_2m_v2(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1148,10 +1137,83 @@ static sd::Tensor<float> sample_lcm(denoise_cb_t model,
                                     sd::Tensor<float> x,
                                     const std::vector<float>& sigmas,
                                     std::shared_ptr<RNG> rng,
-                                    bool is_flow_denoiser) {
+                                    bool is_flow_denoiser,
+                                    const char* extra_sample_args = nullptr) {
+    struct LCMSampleArgs {
+        float noise_clip_std    = 0.0f;
+        float noise_scale_start = 1.0f;
+        float noise_scale_end   = 1.0f;
+    };
+
+    auto trim = [](std::string value) -> std::string {
+        const char* whitespace = " \t\r\n";
+        size_t begin           = value.find_first_not_of(whitespace);
+        if (begin == std::string::npos) {
+            return "";
+        }
+        size_t end = value.find_last_not_of(whitespace);
+        return value.substr(begin, end - begin + 1);
+    };
+
+    LCMSampleArgs args;
+    if (extra_sample_args != nullptr && extra_sample_args[0] != '\0') {
+        std::string raw(extra_sample_args);
+        size_t start                   = 0;
+        bool noise_scale_end_was_set   = false;
+        bool noise_scale_start_was_set = false;
+        auto parse_arg                 = [&](const std::string& item) {
+            std::string token = trim(item);
+            if (token.empty()) {
+                return;
+            }
+            size_t eq = token.find('=');
+            if (eq == std::string::npos) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
+                return;
+            }
+
+            std::string key   = trim(token.substr(0, eq));
+            std::string value = trim(token.substr(eq + 1));
+            float parsed      = 0.0f;
+            try {
+                size_t consumed = 0;
+                parsed          = std::stof(value, &consumed);
+                if (trim(value.substr(consumed)).size() != 0) {
+                    LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
+                    return;
+                }
+            } catch (const std::exception&) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
+                return;
+            }
+
+            if (key == "noise_clip_std") {
+                args.noise_clip_std = parsed;
+            } else if (key == "noise_scale_start") {
+                args.noise_scale_start    = parsed;
+                noise_scale_start_was_set = true;
+            } else if (key == "noise_scale_end") {
+                args.noise_scale_end    = parsed;
+                noise_scale_end_was_set = true;
+            } else {
+                LOG_WARN("ignoring unknown lcm extra sample arg '%s'", key.c_str());
+            }
+        };
+
+        for (size_t pos = 0; pos <= raw.size(); ++pos) {
+            if (pos == raw.size() || raw[pos] == ',' || raw[pos] == ';') {
+                parse_arg(raw.substr(start, pos - start));
+                start = pos + 1;
+            }
+        }
+        if (noise_scale_start_was_set && !noise_scale_end_was_set) {
+            args.noise_scale_end = args.noise_scale_start;
+        }
+    }
+
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1160,7 +1222,27 @@ static sd::Tensor<float> sample_lcm(denoise_cb_t model,
             if (is_flow_denoiser) {
                 x *= (1 - sigmas[i + 1]);
             }
-            x += sd::Tensor<float>::randn_like(x, rng) * sigmas[i + 1];
+            auto noise = sd::Tensor<float>::randn_like(x, rng);
+            if (args.noise_clip_std > 0.0f && noise.numel() > 0) {
+                double mean = 0.0;
+                for (int64_t j = 0; j < noise.numel(); ++j) {
+                    mean += static_cast<double>(noise[j]);
+                }
+                mean /= static_cast<double>(noise.numel());
+
+                double variance = 0.0;
+                for (int64_t j = 0; j < noise.numel(); ++j) {
+                    double centered = static_cast<double>(noise[j]) - mean;
+                    variance += centered * centered;
+                }
+                variance /= static_cast<double>(noise.numel());
+
+                float clip_val = args.noise_clip_std * static_cast<float>(std::sqrt(variance));
+                noise          = sd::ops::clamp(noise, -clip_val, clip_val);
+            }
+            float t           = steps > 1 ? static_cast<float>(i) / static_cast<float>(steps - 1) : 0.0f;
+            float noise_scale = args.noise_scale_start + (args.noise_scale_end - args.noise_scale_start) * t;
+            x += noise * (sigmas[i + 1] * noise_scale);
         }
     }
     return x;
@@ -1177,7 +1259,7 @@ static sd::Tensor<float> sample_ipndm(denoise_cb_t model,
         float sigma      = sigmas[i];
         float sigma_next = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma, i + 1);
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1221,7 +1303,7 @@ static sd::Tensor<float> sample_ipndm_v(denoise_cb_t model,
         float sigma  = sigmas[i];
         float t_next = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma, i + 1);
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1283,7 +1365,7 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1360,7 +1442,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
         float sigma_from = sigmas[i];
         float sigma_to   = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma_from, -(i + 1));
+        auto denoised_opt = model(x, sigma_from, -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1386,7 +1468,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
             sd::Tensor<float> eps1 = denoised - x0;
             sd::Tensor<float> x2   = x0 + eps1 * (h * a21);
 
-            auto denoised2_opt = model(x2, sigma_c2, i + 1);
+            auto denoised2_opt = model(x2, sigma_c2, i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -1463,7 +1545,7 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1);
+        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1, nullptr);
         if (denoised.empty()) {
             return {};
         }
@@ -1539,46 +1621,6 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
     return x;
 }
 
-static sd::Tensor<float> sample_ddim_trailing(denoise_cb_t model,
-                                              sd::Tensor<float> x,
-                                              const std::vector<float>& sigmas,
-                                              std::shared_ptr<RNG> rng,
-                                              float eta) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma    = sigmas[i];
-        float sigma_to = sigmas[i + 1];
-
-        auto model_output_opt = model(x, sigma, i + 1);
-        if (model_output_opt.empty()) {
-            return {};
-        }
-        sd::Tensor<float> model_output = std::move(model_output_opt);
-        model_output                   = (x - model_output) * (1.0f / sigma);
-
-        float alpha_prod_t      = 1.0f / (sigma * sigma + 1.0f);
-        float alpha_prod_t_prev = 1.0f / (sigma_to * sigma_to + 1.0f);
-        float beta_prod_t       = 1.0f - alpha_prod_t;
-
-        sd::Tensor<float> pred_original_sample = ((x / std::sqrt(sigma * sigma + 1)) -
-                                                  std::sqrt(beta_prod_t) * model_output) *
-                                                 (1.0f / std::sqrt(alpha_prod_t));
-
-        float beta_prod_t_prev = 1.0f - alpha_prod_t_prev;
-        float variance         = (beta_prod_t_prev / beta_prod_t) *
-                         (1.0f - alpha_prod_t / alpha_prod_t_prev);
-        float std_dev_t = eta * std::sqrt(variance);
-
-        x = pred_original_sample +
-            std::sqrt((1.0f - alpha_prod_t_prev - std::pow(std_dev_t, 2)) / alpha_prod_t_prev) * model_output;
-
-        if (eta > 0) {
-            x += std_dev_t / std::sqrt(alpha_prod_t_prev) * sd::Tensor<float>::randn_like(x, rng);
-        }
-    }
-    return x;
-}
-
 static sd::Tensor<float> sample_tcd(denoise_cb_t model,
                                     sd::Tensor<float> x,
                                     const std::vector<float>& sigmas,
@@ -1621,12 +1663,12 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
         int timestep_s    = (int)floor((1 - eta) * prev_timestep);
         float sigma       = sigmas[i];
 
-        auto model_output_opt = model(x, sigma, i + 1);
-        if (model_output_opt.empty()) {
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        if (denoised_opt.empty()) {
             return {};
         }
-        sd::Tensor<float> model_output = std::move(model_output_opt);
-        model_output                   = (x - model_output) * (1.0f / sigma);
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> d        = (x - denoised) / sigma;
 
         float alpha_prod_t      = 1.0f / (sigma * sigma + 1.0f);
         float beta_prod_t       = 1.0f - alpha_prod_t;
@@ -1634,12 +1676,8 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
         float alpha_prod_s      = static_cast<float>(alphas_cumprod[timestep_s]);
         float beta_prod_s       = 1.0f - alpha_prod_s;
 
-        sd::Tensor<float> pred_original_sample = ((x / std::sqrt(sigma * sigma + 1)) -
-                                                  std::sqrt(beta_prod_t) * model_output) *
-                                                 (1.0f / std::sqrt(alpha_prod_t));
-
-        x = std::sqrt(alpha_prod_s / alpha_prod_t_prev) * pred_original_sample +
-            std::sqrt(beta_prod_s / alpha_prod_t_prev) * model_output;
+        x = std::sqrt(alpha_prod_s / alpha_prod_t_prev) * denoised +
+            std::sqrt(beta_prod_s / alpha_prod_t_prev) * d;
 
         if (eta > 0 && sigma_to > 0.0f) {
             x = std::sqrt(alpha_prod_t_prev / alpha_prod_s) * x +
@@ -1649,6 +1687,56 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
     return x;
 }
 
+static sd::Tensor<float> sample_euler_cfg_pp(denoise_cb_t model,
+                                             sd::Tensor<float> x,
+                                             const std::vector<float>& sigmas) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma = sigmas[i];
+        sd::Tensor<float> uncond_denoised;
+
+        auto denoised_opt = model(x, sigma, i + 1, &uncond_denoised);
+        if (denoised_opt.empty() || uncond_denoised.empty()) {
+            return {};
+        }
+
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> d        = (x - uncond_denoised) / sigma;
+
+        x = denoised + d * sigmas[i + 1];
+    }
+    return x;
+}
+
+static sd::Tensor<float> sample_euler_ancestral_cfg_pp(denoise_cb_t model,
+                                                       sd::Tensor<float> x,
+                                                       const std::vector<float>& sigmas,
+                                                       std::shared_ptr<RNG> rng,
+                                                       float eta) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma = sigmas[i];
+        sd::Tensor<float> uncond_denoised;
+
+        auto denoised_opt = model(x, sigma, i + 1, &uncond_denoised);
+        if (denoised_opt.empty() || uncond_denoised.empty()) {
+            return {};
+        }
+
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> d        = (x - uncond_denoised) / sigma;
+
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
+
+        x = denoised + d * sigma_down;
+
+        if (sigmas[i + 1] > 0) {
+            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+        }
+    }
+    return x;
+}
+
 // k diffusion reverse ODE: dx = (x - D(x;\sigma)) / \sigma dt; \sigma(t) = t
 static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                             denoise_cb_t model,
@@ -1656,13 +1744,11 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                             std::vector<float> sigmas,
                                             std::shared_ptr<RNG> rng,
                                             float eta,
-                                            bool is_flow_denoiser) {
+                                            bool is_flow_denoiser,
+                                            const char* extra_sample_args) {
     switch (method) {
         case EULER_A_SAMPLE_METHOD:
-            if (is_flow_denoiser)
-                return sample_euler_flow(model, std::move(x), sigmas, rng, eta);
-            else
-                return sample_euler_ancestral(model, std::move(x), sigmas, rng, eta);
+            return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case EULER_SAMPLE_METHOD:
             return sample_euler(model, std::move(x), sigmas);
         case HEUN_SAMPLE_METHOD:
@@ -1679,7 +1765,7 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case DPMPP2Mv2_SAMPLE_METHOD:
             return sample_dpmpp_2m_v2(model, std::move(x), sigmas);
         case LCM_SAMPLE_METHOD:
-            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser);
+            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser, extra_sample_args);
         case IPNDM_SAMPLE_METHOD:
             return sample_ipndm(model, std::move(x), sigmas);
         case IPNDM_V_SAMPLE_METHOD:
@@ -1691,9 +1777,14 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case ER_SDE_SAMPLE_METHOD:
             return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case DDIM_TRAILING_SAMPLE_METHOD:
-            return sample_ddim_trailing(model, std::move(x), sigmas, rng, eta);
+            // DDIM is equivalent to Euler Ancestral with the Simple scheduler
+            return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case TCD_SAMPLE_METHOD:
             return sample_tcd(model, std::move(x), sigmas, rng, eta);
+        case EULER_CFG_PP_SAMPLE_METHOD:
+            return sample_euler_cfg_pp(model, std::move(x), sigmas);
+        case EULER_A_CFG_PP_SAMPLE_METHOD:
+            return sample_euler_ancestral_cfg_pp(model, std::move(x), sigmas, rng, eta);
         default:
             return {};
     }

src/diffusion_model.hpp

@@ -5,6 +5,7 @@
 #include "anima.hpp"
 #include "ernie_image.hpp"
 #include "flux.hpp"
+#include "hidream_o1.hpp"
 #include "ltxv.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
@@ -14,25 +15,31 @@
 #include "z_image.hpp"
 
 struct DiffusionParams {
-    const sd::Tensor<float>* x                        = nullptr;
-    const sd::Tensor<float>* audio_x                  = nullptr;
-    const sd::Tensor<float>* timesteps                = nullptr;
-    const sd::Tensor<float>* audio_timesteps          = nullptr;
-    const sd::Tensor<float>* context                  = nullptr;
-    const sd::Tensor<float>* c_concat                 = nullptr;
-    const sd::Tensor<float>* y                        = nullptr;
-    const sd::Tensor<int32_t>* t5_ids                 = nullptr;
-    const sd::Tensor<float>* t5_weights               = nullptr;
-    const sd::Tensor<float>* guidance                 = nullptr;
-    const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
-    bool increase_ref_index                           = false;
-    int num_video_frames                              = -1;
-    const std::vector<sd::Tensor<float>>* controls    = nullptr;
-    float control_strength                            = 0.f;
-    const sd::Tensor<float>* vace_context             = nullptr;
-    float vace_strength                               = 1.f;
-    int audio_length                                  = 0;
-    const std::vector<int>* skip_layers               = nullptr;
+    const sd::Tensor<float>* x                                         = nullptr;
+    const sd::Tensor<float>* timesteps                                 = nullptr;
+    const sd::Tensor<float>* audio_x                                   = nullptr;
+    const sd::Tensor<float>* audio_timesteps                           = nullptr;
+    const sd::Tensor<float>* context                                   = nullptr;
+    const sd::Tensor<float>* c_concat                                  = nullptr;
+    const sd::Tensor<float>* y                                         = nullptr;
+    const sd::Tensor<int32_t>* t5_ids                                  = nullptr;
+    const sd::Tensor<float>* t5_weights                                = nullptr;
+    const sd::Tensor<float>* guidance                                  = nullptr;
+    const std::vector<sd::Tensor<float>>* ref_latents                  = nullptr;
+    const sd::Tensor<int32_t>* input_ids                               = nullptr;
+    const sd::Tensor<int32_t>* input_pos                               = nullptr;
+    const sd::Tensor<int32_t>* token_types                             = nullptr;
+    const sd::Tensor<int32_t>* vinput_mask                             = nullptr;
+    const std::vector<sd::Tensor<float>>* vlm_images                   = nullptr;
+    const std::vector<std::pair<int, sd::Tensor<float>>>* image_embeds = nullptr;
+    bool increase_ref_index                                            = false;
+    int num_video_frames                                               = -1;
+    const std::vector<sd::Tensor<float>>* controls                     = nullptr;
+    float control_strength                                             = 0.f;
+    const sd::Tensor<float>* vace_context                              = nullptr;
+    float vace_strength                                                = 1.f;
+    int audio_length                                                   = 0;
+    const std::vector<int>* skip_layers                                = nullptr;
 };
 
 template <typename T>
@@ -480,6 +487,82 @@ struct QwenImageModel : public DiffusionModel {
     }
 };
 
+struct HiDreamO1Model : public DiffusionModel {
+    std::string prefix;
+    HiDreamO1::HiDreamO1Runner hidream_o1;
+
+    HiDreamO1Model(ggml_backend_t backend,
+                   bool offload_params_to_cpu,
+                   const String2TensorStorage& tensor_storage_map = {},
+                   const std::string& prefix                      = "model")
+        : prefix(prefix), hidream_o1(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
+    }
+
+    std::string get_desc() override {
+        return hidream_o1.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        hidream_o1.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        hidream_o1.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        hidream_o1.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        hidream_o1.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return hidream_o1.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        hidream_o1.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 0;
+    }
+
+    void set_flash_attention_enabled(bool enabled) {
+        hidream_o1.set_flash_attention_enabled(enabled);
+    }
+
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        hidream_o1.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        hidream_o1.set_circular_axes(circular_x, circular_y);
+    }
+
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        GGML_ASSERT(diffusion_params.input_ids != nullptr);
+        GGML_ASSERT(diffusion_params.input_pos != nullptr);
+        GGML_ASSERT(diffusion_params.token_types != nullptr);
+        static const std::vector<sd::Tensor<float>> empty_images;
+        static const std::vector<std::pair<int, sd::Tensor<float>>> empty_image_embeds;
+        return hidream_o1.compute(n_threads,
+                                  *diffusion_params.x,
+                                  *diffusion_params.timesteps,
+                                  *diffusion_params.input_ids,
+                                  *diffusion_params.input_pos,
+                                  *diffusion_params.token_types,
+                                  tensor_or_empty(diffusion_params.vinput_mask),
+                                  diffusion_params.image_embeds ? *diffusion_params.image_embeds : empty_image_embeds,
+                                  diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_images);
+    }
+};
+
 struct ZImageModel : public DiffusionModel {
     std::string prefix;
     ZImage::ZImageRunner z_image;

src/ggml_extend.hpp

@@ -280,6 +280,9 @@ __STATIC_INLINE__ void print_sd_tensor(const sd::Tensor<T>& tensor, bool shape_o
     if (shape_only) {
         return;
     }
+    if (tensor.empty()) {
+        return;
+    }
     int range                  = 3;
     std::vector<int64_t> shape = tensor.shape();
     while (shape.size() < 4) {
@@ -1713,12 +1716,9 @@ struct GGMLRunnerContext {
         if (debug_tensors == nullptr || tensor == nullptr) {
             return;
         }
-        ggml_tensor* snapshot = tensor;
-        if (!ggml_is_contiguous(snapshot) || snapshot->view_src != nullptr) {
-            snapshot = ggml_cont(ggml_ctx, snapshot);
-        }
-        ggml_tensor* dst = ggml_dup_tensor(ggml_ctx, snapshot);
-        snapshot         = ggml_cpy(ggml_ctx, snapshot, dst);
+        ggml_tensor* snapshot = ggml_cont(ggml_ctx, tensor);
+        ggml_tensor* dst      = ggml_dup_tensor(ggml_ctx, snapshot);
+        snapshot              = ggml_cpy(ggml_ctx, snapshot, dst);
         ggml_set_output(snapshot);
         debug_tensors->push_back({snapshot, name});
     }
@@ -2024,9 +2024,13 @@ protected:
             ggml_backend_buffer_t src_buf = sd::ggml_graph_cut::tensor_buffer(src);
             ggml_backend_buffer_t dst_buf = sd::ggml_graph_cut::tensor_buffer(dst);
             if (src_buf == nullptr || dst_buf == nullptr) {
-                LOG_ERROR("%s cache copy tensor buffer missing: name=%s src_buffer=%p src_view_src=%p src_view_src_buffer=%p dst_buffer=%p",
+                LOG_ERROR("%s cache copy tensor buffer missing: name=%s op=%s src0=%p src0_name=%s src0_buffer=%p src_buffer=%p src_view_src=%p src_view_src_buffer=%p dst_buffer=%p",
                           get_desc().c_str(),
                           src && src->name[0] != '\0' ? src->name : "<unnamed>",
+                          src ? ggml_op_name(src->op) : "<null>",
+                          src ? src->src[0] : nullptr,
+                          (src && src->src[0] && src->src[0]->name[0] != '\0') ? src->src[0]->name : "<unnamed>",
+                          (src && src->src[0]) ? sd::ggml_graph_cut::tensor_buffer(src->src[0]) : nullptr,
                           src ? src->buffer : nullptr,
                           src ? src->view_src : nullptr,
                           (src && src->view_src) ? src->view_src->buffer : nullptr,
@@ -2058,6 +2062,42 @@ protected:
         return true;
     }
 
+    template <typename T>
+    std::optional<sd::Tensor<T>> read_graph_tensor(ggml_tensor* tensor, const char* label) {
+        if (tensor == nullptr) {
+            LOG_ERROR("%s %s tensor is null", get_desc().c_str(), label);
+            return std::nullopt;
+        }
+        if (tensor->type != sd::GGMLTypeTraits<T>::type) {
+            LOG_ERROR("%s %s tensor type mismatch: got %s",
+                      get_desc().c_str(),
+                      label,
+                      ggml_type_name(tensor->type));
+            return std::nullopt;
+        }
+        ggml_backend_buffer_t buf = sd::ggml_graph_cut::tensor_buffer(tensor);
+        if (buf == nullptr) {
+            LOG_ERROR("%s %s tensor buffer missing: name=%s op=%s buffer=%p view_src=%p view_src_buffer=%p data=%p",
+                      get_desc().c_str(),
+                      label,
+                      tensor->name[0] != '\0' ? tensor->name : "<unnamed>",
+                      ggml_op_name(tensor->op),
+                      tensor->buffer,
+                      tensor->view_src,
+                      tensor->view_src ? tensor->view_src->buffer : nullptr,
+                      tensor->data);
+            return std::nullopt;
+        }
+
+        sd::Tensor<T> result(sd::shape_from_ggml(tensor));
+        if (tensor->view_src != nullptr || !ggml_is_contiguous(tensor) || tensor->buffer == nullptr) {
+            ggml_backend_tensor_get(tensor, result.data(), 0, ggml_nbytes(tensor));
+        } else {
+            ggml_backend_tensor_get(tensor, result.data(), 0, ggml_nbytes(tensor));
+        }
+        return result;
+    }
+
     void copy_data_to_backend_tensor(ggml_cgraph* gf, bool clear_after_copy = true) {
         GGML_ASSERT(gf != nullptr);
         std::unordered_set<const ggml_tensor*> graph_tensor_set;
@@ -2078,6 +2118,9 @@ protected:
                 continue;
             }
             const char* name = ggml_get_name(tensor);
+            if (graph_tensor_set.find(tensor) == graph_tensor_set.end()) {
+                continue;
+            }
             if (tensor->buffer == nullptr) {
                 LOG_WARN("%s skip backend tensor copy: tensor buffer not set, name='%s', ne=[%lld,%lld,%lld,%lld], type=%s",
                          get_desc().c_str(),
@@ -2090,10 +2133,6 @@ protected:
                 continue;
             }
 
-            if (graph_tensor_set.find(tensor) == graph_tensor_set.end()) {
-                continue;
-            }
-
             ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
             if (buf == nullptr) {
                 LOG_WARN("%s graph exec skip tensor copy: name=%s op=%s reason=buffer_not_set data=%p view_src=%p view_src_buffer=%p",
@@ -2479,11 +2518,32 @@ protected:
             return std::nullopt;
         }
 
+        std::unordered_set<const ggml_tensor*> debug_graph_tensor_set;
+        const int n_debug_leafs = sd::ggml_graph_cut::leaf_count(gf);
+        const int n_debug_nodes = ggml_graph_n_nodes(gf);
+        debug_graph_tensor_set.reserve(static_cast<size_t>(n_debug_leafs + n_debug_nodes));
+        for (int i = 0; i < n_debug_leafs; ++i) {
+            debug_graph_tensor_set.insert(sd::ggml_graph_cut::leaf_tensor(gf, i));
+        }
+        for (int i = 0; i < n_debug_nodes; ++i) {
+            debug_graph_tensor_set.insert(ggml_graph_node(gf, i));
+        }
+
         for (const auto& entry : debug_tensors) {
             auto tensor = entry.first;
             if (tensor == nullptr) {
                 continue;
             }
+            if (debug_graph_tensor_set.find(tensor) == debug_graph_tensor_set.end()) {
+                continue;
+            }
+            ggml_backend_buffer_t tensor_buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+            if (tensor_buf == nullptr) {
+                LOG_WARN("%s skip debug tensor '%s': tensor buffer not set",
+                         get_desc().c_str(),
+                         entry.second.c_str());
+                continue;
+            }
             if (tensor->type != GGML_TYPE_F32) {
                 LOG_WARN("%s skip debug tensor '%s': only GGML_TYPE_F32 is supported, got %s",
                          get_desc().c_str(),
@@ -2508,7 +2568,15 @@ protected:
         auto result         = ggml_get_tensor(compute_ctx, final_result_name.c_str());
         std::optional<sd::Tensor<T>> output;
         if (!no_return) {
-            output = sd::make_sd_tensor_from_ggml<T>(result);
+            output = read_graph_tensor<T>(result, "output");
+            if (!output.has_value()) {
+                if (free_compute_buffer_immediately) {
+                    free_compute_buffer();
+                } else if (use_partial_param_offload) {
+                    restore_partial_params();
+                }
+                return std::nullopt;
+            }
         } else {
             output = sd::Tensor<T>();
         }
@@ -2648,7 +2716,27 @@ public:
 
     bool alloc_params_buffer() {
         size_t num_tensors = ggml_tensor_num(params_ctx);
-        params_buffer      = ggml_backend_alloc_ctx_tensors(params_ctx, params_backend);
+        if (num_tensors > 0) {
+            // ggml_backend_alloc_ctx_tensors fails when all tensors are already allocated
+            // (typical for memory-mapped weights). See ggml-alloc.c n_buffers==0 branch.
+            bool all_have_data = true;
+            for (ggml_tensor* t = ggml_get_first_tensor(params_ctx); t != nullptr; t = ggml_get_next_tensor(params_ctx, t)) {
+                if (t->data == nullptr) {
+                    all_have_data = false;
+                    break;
+                }
+            }
+            if (all_have_data) {
+                LOG_DEBUG("%s all params already mmap-allocated (no separate buffer needed)", get_desc().c_str());
+                params_buffer = nullptr;
+                rebuild_params_tensor_set();
+                return true;
+            }
+        } else {
+            LOG_DEBUG("%s skipping params allocation (no tensors)", get_desc().c_str());
+            return true;
+        }
+        params_buffer = ggml_backend_alloc_ctx_tensors(params_ctx, params_backend);
         if (params_buffer == nullptr) {
             LOG_ERROR("%s alloc params backend buffer failed, num_tensors = %i",
                       get_desc().c_str(),

src/ggml_graph_cut.cpp

@@ -16,6 +16,9 @@
 
 namespace sd::ggml_graph_cut {
 
+    static constexpr double MAX_VRAM_BYTES_PER_GIB      = 1024.0 * 1024.0 * 1024.0;
+    static constexpr size_t MAX_VRAM_AUTO_RESERVE_BYTES = 1024ULL * 1024ULL * 1024ULL;
+
     static std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
         if (tensor == nullptr) {
             return "<null>";
@@ -45,6 +48,21 @@ namespace sd::ggml_graph_cut {
         return params_tensor_set.find(tensor) != params_tensor_set.end();
     }
 
+    static int graph_node_index_by_name(ggml_cgraph* gf, const char* name) {
+        GGML_ASSERT(gf != nullptr);
+        if (name == nullptr || name[0] == '\0') {
+            return -1;
+        }
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            if (node != nullptr && std::strcmp(node->name, name) == 0) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
     static Plan::InputShape input_shape(const ggml_tensor* tensor) {
         Plan::InputShape shape;
         if (tensor == nullptr) {
@@ -64,6 +82,58 @@ namespace sd::ggml_graph_cut {
                segment.output_bytes;
     }
 
+    size_t max_vram_gib_to_bytes(float max_vram) {
+        if (max_vram <= 0.f) {
+            return 0;
+        }
+        return static_cast<size_t>(static_cast<double>(max_vram) * MAX_VRAM_BYTES_PER_GIB);
+    }
+
+    static float max_vram_bytes_to_gib(size_t max_vram_bytes) {
+        return static_cast<float>(static_cast<double>(max_vram_bytes) / MAX_VRAM_BYTES_PER_GIB);
+    }
+
+    static size_t resolve_auto_max_vram_bytes(ggml_backend_t backend) {
+        if (backend == nullptr) {
+            LOG_WARN("--max-vram -1 requested, but no backend is available; disabling graph splitting");
+            return 0;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        if (dev == nullptr) {
+            LOG_WARN("--max-vram -1 requested, but no backend device is available; disabling graph splitting");
+            return 0;
+        }
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
+            LOG_WARN("--max-vram -1 requested, but the main backend is CPU; disabling graph splitting");
+            return 0;
+        }
+
+        size_t free_vram  = 0;
+        size_t total_vram = 0;
+        ggml_backend_dev_memory(dev, &free_vram, &total_vram);
+
+        if (free_vram <= MAX_VRAM_AUTO_RESERVE_BYTES) {
+            LOG_WARN("--max-vram -1 requested, but free VRAM is %.2f GiB; reserving 1.00 GiB leaves no graph budget",
+                     free_vram / MAX_VRAM_BYTES_PER_GIB);
+            return 0;
+        }
+
+        const size_t max_vram_bytes = free_vram - MAX_VRAM_AUTO_RESERVE_BYTES;
+        LOG_INFO("--max-vram -1 auto-detected %.2f GiB free VRAM (%.2f GiB total), reserving 1.00 GiB; using %.2f GiB",
+                 free_vram / MAX_VRAM_BYTES_PER_GIB,
+                 total_vram / MAX_VRAM_BYTES_PER_GIB,
+                 max_vram_bytes / MAX_VRAM_BYTES_PER_GIB);
+        return max_vram_bytes;
+    }
+
+    float resolve_max_vram_gib(float max_vram, ggml_backend_t backend) {
+        if (max_vram != -1.f) {
+            return max_vram;
+        }
+        return max_vram_bytes_to_gib(resolve_auto_max_vram_bytes(backend));
+    }
+
     static Segment make_segment_seed(const Plan& plan,
                                      size_t start_segment_index,
                                      size_t end_segment_index) {
@@ -244,6 +314,11 @@ namespace sd::ggml_graph_cut {
         if (tensor == nullptr) {
             return nullptr;
         }
+        if (tensor_buffer(tensor) == nullptr && tensor->src[0] != nullptr &&
+            ggml_nelements(tensor->src[0]) == ggml_nelements(tensor) &&
+            ggml_nbytes(tensor->src[0]) == ggml_nbytes(tensor)) {
+            return cache_source_tensor(tensor->src[0]);
+        }
         return tensor->view_src ? tensor->view_src : tensor;
     }
 
@@ -503,11 +578,15 @@ namespace sd::ggml_graph_cut {
                           log_desc);
         }
 
-        ggml_tensor* final_output = ggml_graph_node(gf, -1);
-        if (final_output != nullptr && available_cut_output_node_indices.find(n_nodes - 1) == available_cut_output_node_indices.end()) {
+        int final_output_index = graph_node_index_by_name(gf, "ggml_runner_final_result_tensor");
+        if (final_output_index < 0) {
+            final_output_index = n_nodes - 1;
+        }
+        ggml_tensor* final_output = final_output_index >= 0 ? ggml_graph_node(gf, final_output_index) : nullptr;
+        if (final_output != nullptr && available_cut_output_node_indices.find(final_output_index) == available_cut_output_node_indices.end()) {
             Segment final_segment;
             final_segment.group_name = "ggml_runner.final";
-            final_segment.output_node_indices.push_back(n_nodes - 1);
+            final_segment.output_node_indices.push_back(final_output_index);
             build_segment(gf,
                           plan,
                           final_segment,

src/ggml_graph_cut.h

@@ -83,6 +83,8 @@ namespace sd::ggml_graph_cut {
                                           ggml_cgraph* gf,
                                           const Segment& segment,
                                           const char* log_desc);
+    size_t max_vram_gib_to_bytes(float max_vram);
+    float resolve_max_vram_gib(float max_vram, ggml_backend_t backend);
     Plan build_plan(ggml_backend_t backend,
                     ggml_cgraph* gf,
                     const std::unordered_set<const ggml_tensor*>& params_tensor_set,

src/hidream_o1.hpp

@@ -0,0 +1,653 @@
+#ifndef __SD_HIDREAM_O1_H__
+#define __SD_HIDREAM_O1_H__
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common_dit.hpp"
+#include "conditioner.hpp"
+#include "llm.hpp"
+#include "util.h"
+
+namespace HiDreamO1 {
+    constexpr int HIDREAM_O1_GRAPH_SIZE = 32768;
+    constexpr int PATCH_SIZE            = 32;
+    constexpr int TIMESTEP_TOKEN_NUM    = 1;
+    constexpr int IMAGE_TOKEN_ID        = 151655;
+    constexpr int VISION_START_TOKEN_ID = 151652;
+
+    static inline std::string repeat_special_token(const std::string& token, int64_t count) {
+        std::string out;
+        out.reserve(static_cast<size_t>(count) * token.size());
+        for (int64_t i = 0; i < count; ++i) {
+            out += token;
+        }
+        return out;
+    }
+
+    static inline std::pair<int, int> calculate_dimensions(int max_size, double ratio) {
+        int width  = static_cast<int>(std::sqrt(max_size * max_size * ratio));
+        int height = static_cast<int>(width / ratio);
+        width      = (width / PATCH_SIZE) * PATCH_SIZE;
+        height     = (height / PATCH_SIZE) * PATCH_SIZE;
+        width      = std::max(width, PATCH_SIZE);
+        height     = std::max(height, PATCH_SIZE);
+        return {width, height};
+    }
+
+    static inline sd::Tensor<float> resize_to_area(const sd::Tensor<float>& image, int image_size) {
+        int64_t width  = image.shape()[0];
+        int64_t height = image.shape()[1];
+        int64_t s_max  = static_cast<int64_t>(image_size) * image_size;
+        double scale   = std::sqrt(static_cast<double>(s_max) / static_cast<double>(width * height));
+
+        std::vector<std::pair<int64_t, int64_t>> sizes = {
+            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+        };
+        std::sort(sizes.begin(), sizes.end(), [](const auto& a, const auto& b) {
+            return a.first * a.second > b.first * b.second;
+        });
+
+        std::pair<int64_t, int64_t> new_size = sizes.back();
+        for (const auto& size : sizes) {
+            if (size.first > 0 && size.second > 0 && size.first * size.second <= s_max) {
+                new_size = size;
+                break;
+            }
+        }
+
+        double s1 = static_cast<double>(width) / static_cast<double>(new_size.first);
+        double s2 = static_cast<double>(height) / static_cast<double>(new_size.second);
+        sd::Tensor<float> resized;
+        if (s1 < s2) {
+            int64_t resized_h = static_cast<int64_t>(std::llround(height / s1));
+            resized           = sd::ops::interpolate(image,
+                                                     {new_size.first, resized_h, image.shape()[2], image.shape()[3]},
+                                                     sd::ops::InterpolateMode::Bicubic);
+            int64_t top       = (resized_h - new_size.second) / 2;
+            resized           = sd::ops::slice(resized, 1, top, top + new_size.second);
+        } else {
+            int64_t resized_w = static_cast<int64_t>(std::llround(width / s2));
+            resized           = sd::ops::interpolate(image,
+                                                     {resized_w, new_size.second, image.shape()[2], image.shape()[3]},
+                                                     sd::ops::InterpolateMode::Bicubic);
+            int64_t left      = (resized_w - new_size.first) / 2;
+            resized           = sd::ops::slice(resized, 0, left, left + new_size.first);
+        }
+        return resized;
+    }
+
+    static inline std::vector<int32_t> build_position_ids(const std::vector<int32_t>& input_ids,
+                                                          const std::vector<std::array<int32_t, 3>>& image_grids,
+                                                          const std::vector<int32_t>& skip_vision_start_token) {
+        std::vector<int32_t> position_ids(4 * input_ids.size(), 0);
+        int image_index = 0;
+        int st          = 0;
+        int fix_point   = 4096;
+        std::vector<int32_t> out_t;
+        std::vector<int32_t> out_h;
+        std::vector<int32_t> out_w;
+
+        while (st < static_cast<int>(input_ids.size())) {
+            int ed = st;
+            while (ed < static_cast<int>(input_ids.size()) && input_ids[ed] != IMAGE_TOKEN_ID) {
+                ed++;
+            }
+
+            if (ed >= static_cast<int>(input_ids.size())) {
+                int st_idx = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
+                for (int i = 0; i < static_cast<int>(input_ids.size()) - st; ++i) {
+                    out_t.push_back(st_idx + i);
+                    out_h.push_back(st_idx + i);
+                    out_w.push_back(st_idx + i);
+                }
+                break;
+            }
+
+            int text_len = std::max(0, ed - st - skip_vision_start_token[image_index]);
+            int st_idx   = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
+            for (int i = 0; i < text_len; ++i) {
+                out_t.push_back(st_idx + i);
+                out_h.push_back(st_idx + i);
+                out_w.push_back(st_idx + i);
+            }
+
+            auto grid = image_grids[image_index];
+            int base;
+            if (skip_vision_start_token[image_index]) {
+                if (fix_point > 0) {
+                    base      = fix_point;
+                    fix_point = 0;
+                } else {
+                    base = st_idx;
+                }
+            } else {
+                base = text_len + st_idx;
+            }
+            for (int32_t ti = 0; ti < grid[0]; ++ti) {
+                for (int32_t hi = 0; hi < grid[1]; ++hi) {
+                    for (int32_t wi = 0; wi < grid[2]; ++wi) {
+                        out_t.push_back(base + ti);
+                        out_h.push_back(base + hi);
+                        out_w.push_back(base + wi);
+                    }
+                }
+            }
+
+            st = ed + grid[0] * grid[1] * grid[2];
+            image_index++;
+        }
+
+        GGML_ASSERT(out_t.size() == input_ids.size());
+        for (size_t i = 0; i < input_ids.size(); ++i) {
+            // ggml IMROPE consumes 4 flattened position streams:
+            //   [t, h, w, e]
+            // llama.cpp's generic Qwen-VL fallback expands text positions as
+            // [pos, pos, pos, 0]. Keep the extra stream zeroed here too.
+            position_ids[i]                        = out_t[i];
+            position_ids[input_ids.size() + i]     = out_h[i];
+            position_ids[input_ids.size() * 2 + i] = out_w[i];
+            position_ids[input_ids.size() * 3 + i] = 0;
+        }
+        return position_ids;
+    }
+
+    struct TimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size = 256;
+
+        TimestepEmbedder(int64_t hidden_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
+            auto emb   = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10000, 1000.0f);
+            emb        = mlp_0->forward(ctx, emb);
+            emb        = ggml_silu_inplace(ctx->ggml_ctx, emb);
+            emb        = mlp_2->forward(ctx, emb);
+            return emb;
+        }
+    };
+
+    struct BottleneckPatchEmbed : public GGMLBlock {
+        BottleneckPatchEmbed(int64_t in_dim, int64_t pca_dim, int64_t embed_dim) {
+            blocks["proj1"] = std::make_shared<Linear>(in_dim, pca_dim, false);
+            blocks["proj2"] = std::make_shared<Linear>(pca_dim, embed_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj1 = std::dynamic_pointer_cast<Linear>(blocks["proj1"]);
+            auto proj2 = std::dynamic_pointer_cast<Linear>(blocks["proj2"]);
+            return proj2->forward(ctx, proj1->forward(ctx, x));
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t out_dim) {
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, out_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, x);
+        }
+    };
+
+    struct HiDreamO1Params {
+        LLM::LLMParams llm;
+        int patch_size = PATCH_SIZE;
+    };
+
+    static inline HiDreamO1Params make_hidream_o1_params() {
+        HiDreamO1Params params;
+        params.llm.arch                           = LLM::LLMArch::QWEN3_VL;
+        params.llm.hidden_size                    = 4096;
+        params.llm.intermediate_size              = 12288;
+        params.llm.num_layers                     = 36;
+        params.llm.num_heads                      = 32;
+        params.llm.num_kv_heads                   = 8;
+        params.llm.head_dim                       = 128;
+        params.llm.qkv_bias                       = false;
+        params.llm.qk_norm                        = true;
+        params.llm.vocab_size                     = 151936;
+        params.llm.rms_norm_eps                   = 1e-6f;
+        params.llm.vision.arch                    = LLM::LLMVisionArch::QWEN3_VL;
+        params.llm.vision.num_layers              = 27;
+        params.llm.vision.hidden_size             = 1152;
+        params.llm.vision.intermediate_size       = 4304;
+        params.llm.vision.num_heads               = 16;
+        params.llm.vision.out_hidden_size         = 4096;
+        params.llm.vision.patch_size              = 16;
+        params.llm.vision.spatial_merge_size      = 2;
+        params.llm.vision.temporal_patch_size     = 2;
+        params.llm.vision.num_position_embeddings = 2304;
+        return params;
+    }
+
+    struct HiDreamO1Model : public GGMLBlock {
+        HiDreamO1Params params;
+
+        HiDreamO1Model() = default;
+        explicit HiDreamO1Model(HiDreamO1Params params)
+            : params(std::move(params)) {
+            blocks["language_model"] = std::make_shared<LLM::TextModel>(this->params.llm);
+            blocks["t_embedder1"]    = std::make_shared<TimestepEmbedder>(this->params.llm.hidden_size);
+            blocks["x_embedder"]     = std::make_shared<BottleneckPatchEmbed>(this->params.patch_size * this->params.patch_size * 3,
+                                                                          this->params.llm.hidden_size / 4,
+                                                                          this->params.llm.hidden_size);
+            blocks["final_layer2"]   = std::make_shared<FinalLayer>(this->params.llm.hidden_size,
+                                                                  this->params.patch_size * this->params.patch_size * 3);
+        }
+
+        std::shared_ptr<LLM::TextModel> text_model() {
+            return std::dynamic_pointer_cast<LLM::TextModel>(blocks["language_model"]);
+        }
+
+        std::shared_ptr<TimestepEmbedder> timestep_embedder() {
+            return std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder1"]);
+        }
+
+        std::shared_ptr<BottleneckPatchEmbed> patch_embedder() {
+            return std::dynamic_pointer_cast<BottleneckPatchEmbed>(blocks["x_embedder"]);
+        }
+
+        std::shared_ptr<FinalLayer> final_layer() {
+            return std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer2"]);
+        }
+    };
+
+    struct HiDreamO1VisionRunner : public GGMLRunner {
+        HiDreamO1Params params;
+        std::shared_ptr<LLM::VisionModel> model;
+
+        std::vector<int> window_index_vec;
+        std::vector<int> window_inverse_index_vec;
+        std::vector<float> window_mask_vec;
+        std::vector<float> pe_vec;
+        std::array<std::vector<int32_t>, 4> pos_embed_idx_data_;
+        std::array<std::vector<float>, 4> pos_embed_weight_data_;
+
+        HiDreamO1VisionRunner(ggml_backend_t backend,
+                              bool offload_params_to_cpu,
+                              const String2TensorStorage& tensor_storage_map = {},
+                              const std::string& prefix                      = "model.visual")
+            : GGMLRunner(backend, offload_params_to_cpu),
+              params(make_hidream_o1_params()),
+              model(std::make_shared<LLM::VisionModel>(false, params.llm.vision)) {
+            model->init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "hidream_o1_vision";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix = "model.visual") {
+            model->get_param_tensors(tensors, prefix);
+        }
+
+        ggml_tensor* encode_image(GGMLRunnerContext* runner_ctx, ggml_tensor* image) {
+            return LLM::LLMRunner::encode_image_common(this,
+                                                       compute_ctx,
+                                                       runner_ctx,
+                                                       image,
+                                                       params.llm.vision,
+                                                       model,
+                                                       window_index_vec,
+                                                       window_inverse_index_vec,
+                                                       window_mask_vec,
+                                                       pe_vec,
+                                                       pos_embed_idx_data_,
+                                                       pos_embed_weight_data_);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& image_tensor) {
+            ggml_cgraph* gf    = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
+            ggml_tensor* image = make_input(image_tensor);
+            auto runner_ctx    = get_context();
+            auto image_embeds  = encode_image(&runner_ctx, image);
+            ggml_build_forward_expand(gf, image_embeds);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& image) {
+            auto get_graph = [&]() {
+                return build_graph(image);
+            };
+            auto output = GGMLRunner::compute<float>(get_graph, n_threads, false);
+            return output.has_value() ? std::move(output.value()) : sd::Tensor<float>();
+        }
+    };
+
+    struct HiDreamO1Runner : public GGMLRunner {
+        HiDreamO1Params params;
+        HiDreamO1Model model;
+
+        std::vector<float> attention_mask_vec;
+
+        HiDreamO1Runner(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2TensorStorage& tensor_storage_map = {},
+                        const std::string& prefix                      = "model")
+            : GGMLRunner(backend, offload_params_to_cpu),
+              params(make_hidream_o1_params()) {
+            model = HiDreamO1Model(params);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "hidream_o1";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timestep_tensor,
+                                 const sd::Tensor<int32_t>& input_ids_tensor,
+                                 const sd::Tensor<int32_t>& input_pos_tensor,
+                                 const sd::Tensor<int32_t>& token_types_tensor,
+                                 const sd::Tensor<int32_t>& vinput_mask_tensor,
+                                 const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds_tensor,
+                                 const std::vector<sd::Tensor<float>>& ref_images) {
+            ggml_cgraph* gf        = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timestep  = make_input(timestep_tensor);
+            ggml_tensor* input_ids = make_input(input_ids_tensor);
+            ggml_tensor* input_pos = make_input(input_pos_tensor);
+
+            auto text_model   = model.text_model();
+            auto t_embedder1  = model.timestep_embedder();
+            auto x_embedder   = model.patch_embedder();
+            auto final_layer2 = model.final_layer();
+
+            std::vector<ggml_tensor*> ref_image_tensors;
+            for (const auto& image : ref_images) {
+                ref_image_tensors.push_back(make_input(image));
+            }
+
+            attention_mask_vec    = std::vector<float>(static_cast<size_t>(token_types_tensor.shape()[0] * token_types_tensor.shape()[0]), 0.0f);
+            int64_t total_seq_len = token_types_tensor.shape()[0];
+            for (int64_t query = 0; query < total_seq_len; ++query) {
+                bool is_gen = token_types_tensor.values()[static_cast<size_t>(query)] > 0;
+                for (int64_t key = 0; key < total_seq_len; ++key) {
+                    if (!is_gen && key > query) {
+                        attention_mask_vec[static_cast<size_t>(query * total_seq_len + key)] = -INFINITY;
+                    }
+                }
+            }
+            auto attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, total_seq_len, total_seq_len);
+            set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+
+            auto runner_ctx = get_context();
+            auto txt        = text_model->embed(&runner_ctx, input_ids);
+            std::vector<std::pair<int, ggml_tensor*>> image_embeds;
+            image_embeds.reserve(image_embeds_tensor.size());
+            for (const auto& image_embed : image_embeds_tensor) {
+                image_embeds.emplace_back(image_embed.first, make_input(image_embed.second));
+            }
+            txt = LLM::splice_image_embeds(&runner_ctx, txt, image_embeds);
+
+            auto t_emb          = t_embedder1->forward(&runner_ctx, timestep);
+            int64_t txt_seq_len = input_ids->ne[0];
+            if (txt_seq_len > 1) {
+                auto prefix = ggml_ext_slice(compute_ctx, txt, 1, 0, txt_seq_len - 1);
+                txt         = ggml_concat(compute_ctx, prefix, ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1), 1);
+            } else {
+                txt = ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1);
+            }
+
+            auto vinputs          = DiT::pad_and_patchify(&runner_ctx, x, PATCH_SIZE, PATCH_SIZE);
+            int64_t target_tokens = vinputs->ne[1];
+            for (ggml_tensor* ref_image : ref_image_tensors) {
+                auto ref = DiT::pad_and_patchify(&runner_ctx, ref_image, PATCH_SIZE, PATCH_SIZE);
+                vinputs  = ggml_concat(compute_ctx, vinputs, ref, 1);
+            }
+            auto vis = x_embedder->forward(&runner_ctx, vinputs);
+
+            auto inputs_embeds = ggml_concat(compute_ctx, txt, vis, 1);
+            auto hidden_states = text_model->forward_embeds(&runner_ctx, inputs_embeds, input_pos, attention_mask, {});
+            auto x_pred_all    = final_layer2->forward(&runner_ctx, hidden_states);
+
+            int64_t x_pred_start = txt_seq_len;
+            if (!vinput_mask_tensor.empty()) {
+                int64_t seq_len      = static_cast<int64_t>(vinput_mask_tensor.shape()[0]);
+                int64_t first_vinput = 0;
+                while (first_vinput < seq_len && vinput_mask_tensor.values()[static_cast<size_t>(first_vinput)] == 0) {
+                    first_vinput++;
+                }
+                x_pred_start = first_vinput;
+            }
+            auto x_pred = ggml_ext_slice(compute_ctx, x_pred_all, 1, x_pred_start, x_pred_start + target_tokens);
+            x_pred      = DiT::unpatchify_and_crop(compute_ctx, x_pred, x->ne[1], x->ne[0], PATCH_SIZE, PATCH_SIZE);
+
+            float sigma = 1.0f - timestep_tensor.values()[0];
+            sigma       = std::max(1e-6f, sigma);
+            auto out    = ggml_scale(compute_ctx, ggml_sub(compute_ctx, x, x_pred), 1.0f / sigma);
+
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timestep,
+                                  const sd::Tensor<int32_t>& input_ids,
+                                  const sd::Tensor<int32_t>& input_pos,
+                                  const sd::Tensor<int32_t>& token_types,
+                                  const sd::Tensor<int32_t>& vinput_mask,
+                                  const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
+                                  const std::vector<sd::Tensor<float>>& ref_images) {
+            auto get_graph = [&]() {
+                return build_graph(x, timestep, input_ids, input_pos, token_types, vinput_mask, image_embeds, ref_images);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+    };
+
+    struct HiDreamO1Conditioner : public Conditioner {
+        Qwen2Tokenizer tokenizer;
+        std::shared_ptr<HiDreamO1VisionRunner> vision_runner;
+
+        HiDreamO1Conditioner(ggml_backend_t backend,
+                             bool offload_params_to_cpu,
+                             const String2TensorStorage& tensor_storage_map = {})
+            : vision_runner(std::make_shared<HiDreamO1VisionRunner>(backend, offload_params_to_cpu, tensor_storage_map)) {}
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+            vision_runner->get_param_tensors(tensors);
+        }
+
+        void alloc_params_buffer() override {
+            vision_runner->alloc_params_buffer();
+        }
+
+        void free_params_buffer() override {
+            vision_runner->free_params_buffer();
+        }
+
+        size_t get_params_buffer_size() override {
+            return vision_runner->get_params_buffer_size();
+        }
+
+        void set_max_graph_vram_bytes(size_t max_graph_vram_bytes) override {
+            vision_runner->set_max_graph_vram_bytes(max_graph_vram_bytes);
+        }
+
+        void set_flash_attention_enabled(bool enabled) override {
+            vision_runner->set_flash_attention_enabled(enabled);
+        }
+
+        void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+            vision_runner->set_weight_adapter(adapter);
+        }
+
+        SDCondition get_learned_condition(int n_threads,
+                                          const ConditionerParams& conditioner_params) override {
+            SDCondition result;
+
+            int width                = conditioner_params.width;
+            int height               = conditioner_params.height;
+            int64_t target_image_len = static_cast<int64_t>(width / PATCH_SIZE) * static_cast<int64_t>(height / PATCH_SIZE);
+
+            std::vector<sd::Tensor<float>> ref_images;
+            if (conditioner_params.ref_images != nullptr) {
+                ref_images = *conditioner_params.ref_images;
+            }
+
+            std::vector<std::pair<int, sd::Tensor<float>>> vlm_images;
+            std::vector<std::array<int32_t, 3>> image_grids;
+            std::vector<int32_t> skip_vision_start;
+
+            std::string prompt = "<|im_start|>user\n";
+
+            if (ref_images.empty()) {
+                prompt += conditioner_params.text;
+                prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
+                auto input_ids = tokenizer.encode(prompt, nullptr);
+
+                std::vector<int32_t> input_ids_pad = input_ids;
+                input_ids_pad.push_back(VISION_START_TOKEN_ID);
+                input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
+
+                image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
+                skip_vision_start.push_back(1);
+
+                std::vector<int32_t> token_types(input_ids_pad.size(), 0);
+                int txt_seq_len = static_cast<int>(input_ids.size());
+                int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
+                for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
+                    token_types[i] = 1;
+                }
+
+                auto position_ids = build_position_ids(input_ids_pad, image_grids, skip_vision_start);
+
+                std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
+                std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
+                std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
+                std::vector<int32_t> vinput_mask(token_types.size(), 0);
+                for (int64_t i = txt_seq_len; i < static_cast<int64_t>(vinput_mask.size()); ++i) {
+                    vinput_mask[static_cast<size_t>(i)] = 1;
+                }
+                std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
+
+                result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
+                result.c_position_ids = sd::Tensor<int32_t>(position_shape, position_ids);
+                result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
+                result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
+                return result;
+            }
+
+            int K = static_cast<int>(ref_images.size());
+            int max_size;
+            if (K == 1) {
+                max_size = std::max(height, width);
+            } else if (K == 2) {
+                max_size = std::max(height, width) * 48 / 64;
+            } else if (K <= 4) {
+                max_size = std::max(height, width) / 2;
+            } else if (K <= 8) {
+                max_size = std::max(height, width) * 24 / 64;
+            } else {
+                max_size = std::max(height, width) / 4;
+            }
+
+            int cond_img_size;
+            if (K <= 4) {
+                cond_img_size = 384;
+            } else if (K <= 8) {
+                cond_img_size = 384 * 48 / 64;
+            } else {
+                cond_img_size = 384 / 2;
+            }
+
+            for (const auto& ref_image : ref_images) {
+                auto resized_ref = resize_to_area(ref_image, max_size);
+                resized_ref      = sd::ops::clamp(resized_ref, 0.0f, 1.0f);
+
+                // VLM image: Qwen3-VL expects mean=[0.5]/std=[0.5] (i.e. range [-1,1]),
+                // not CLIP normalization. Resize the already-resized ref directly to
+                // (cond_w, cond_h) to match the Python pipeline's pil_r.resize().
+                auto dims                   = calculate_dimensions(cond_img_size,
+                                                                   static_cast<double>(resized_ref.shape()[0]) / static_cast<double>(resized_ref.shape()[1]));
+                sd::Tensor<float> vlm_image = sd::ops::interpolate(
+                    resized_ref,
+                    {dims.first, dims.second, resized_ref.shape()[2], resized_ref.shape()[3]});
+                vlm_image            = vlm_image * 2.0f - 1.0f;
+                int64_t image_tokens = static_cast<int64_t>(dims.first / PATCH_SIZE) * static_cast<int64_t>(dims.second / PATCH_SIZE);
+
+                auto patch_img = resized_ref * 2.0f - 1.0f;
+                result.c_ref_images.push_back(std::move(patch_img));
+                int64_t prompt_start = static_cast<int64_t>(tokenizer.encode(prompt + "<|vision_start|>", nullptr).size());
+                prompt += "<|vision_start|>";
+                prompt += repeat_special_token("<|image_pad|>", image_tokens);
+                prompt += "<|vision_end|>";
+                vlm_images.emplace_back(static_cast<int>(prompt_start), std::move(vlm_image));
+                image_grids.push_back({1, dims.second / PATCH_SIZE, dims.first / PATCH_SIZE});
+                skip_vision_start.push_back(0);
+            }
+
+            prompt += conditioner_params.text;
+            prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
+            auto input_ids = tokenizer.encode(prompt, nullptr);
+
+            std::vector<int32_t> input_ids_pad = input_ids;
+            input_ids_pad.push_back(VISION_START_TOKEN_ID);
+            input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
+            image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
+            skip_vision_start.push_back(1);
+
+            for (const auto& ref_image : result.c_ref_images) {
+                int64_t ref_len = static_cast<int64_t>(ref_image.shape()[0] / PATCH_SIZE) * static_cast<int64_t>(ref_image.shape()[1] / PATCH_SIZE);
+                input_ids_pad.push_back(VISION_START_TOKEN_ID);
+                input_ids_pad.insert(input_ids_pad.end(), ref_len - 1, IMAGE_TOKEN_ID);
+                image_grids.push_back({1, static_cast<int32_t>(ref_image.shape()[1] / PATCH_SIZE), static_cast<int32_t>(ref_image.shape()[0] / PATCH_SIZE)});
+                skip_vision_start.push_back(1);
+            }
+
+            std::vector<int32_t> token_types(input_ids_pad.size(), 0);
+            int txt_seq_len = static_cast<int>(input_ids.size());
+            int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
+            for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
+                token_types[i] = 1;
+            }
+
+            std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
+            std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
+            std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
+            std::vector<int32_t> vinput_mask(token_types.size(), 0);
+            for (int i = txt_seq_len; i < static_cast<int>(vinput_mask.size()); ++i) {
+                vinput_mask[static_cast<size_t>(i)] = 1;
+            }
+            std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
+
+            result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
+            result.c_position_ids = sd::Tensor<int32_t>(position_shape, build_position_ids(input_ids_pad, image_grids, skip_vision_start));
+            result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
+            result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
+            result.c_image_embeds.reserve(vlm_images.size());
+            for (const auto& vlm_image : vlm_images) {
+                auto image_embed = vision_runner->compute(n_threads, vlm_image.second);
+                if (image_embed.empty()) {
+                    LOG_ERROR("hidream_o1 conditioner: encode VLM image failed");
+                    return SDCondition();
+                }
+                result.c_image_embeds.emplace_back(vlm_image.first, std::move(image_embed));
+            }
+            return result;
+        }
+    };
+}  // namespace HiDreamO1
+
+#endif  // __SD_HIDREAM_O1_H__

src/model.cpp

@@ -437,6 +437,10 @@ SDVersion ModelLoader::get_sd_version() {
         if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
             return VERSION_SD3;
         }
+        if (tensor_storage.name.find("model.x_embedder.proj1.weight") != std::string::npos &&
+            tensor_storage_map.find("model.language_model.layers.0.self_attn.q_proj.weight") != tensor_storage_map.end()) {
+            return VERSION_HIDREAM_O1;
+        }
         if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
             return VERSION_QWEN_IMAGE;
         }
@@ -733,16 +737,10 @@ void ModelLoader::set_wtype_override(ggml_type wtype, std::string tensor_type_ru
     }
 }
 
-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);
-    std::atomic<int64_t> copy_to_backend_time_ms(0);
-    std::atomic<int64_t> convert_time_ms(0);
-    std::atomic<uint64_t> bytes_processed(0);
-
-    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
-    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
+void ModelLoader::process_model_files(bool enable_mmap, bool writable_mmap) {
+    if (model_files_processed) {
+        return;
+    }
 
     int64_t start_time = ggml_time_ms();
 
@@ -754,22 +752,13 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
         processed_tensor_storages.push_back(tensor_storage);
     }
 
-    process_time_ms = ggml_time_ms() - start_time;
-
-    bool success                          = true;
-    size_t total_tensors_processed        = 0;
-    const size_t total_tensors_to_process = processed_tensor_storages.size();
-    const int64_t t_start                 = ggml_time_ms();
-    int last_n_threads                    = 1;
-
     for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
         std::string file_path = file_paths_[file_index];
-        LOG_DEBUG("loading tensors from %s", file_path.c_str());
 
-        std::vector<const TensorStorage*> file_tensors;
+        std::vector<TensorStorage> file_tensors;
         for (const auto& ts : processed_tensor_storages) {
             if (ts.file_index == file_index) {
-                file_tensors.push_back(&ts);
+                file_tensors.push_back(ts);
             }
         }
         if (file_tensors.empty()) {
@@ -778,21 +767,169 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
 
         bool is_zip = false;
         for (auto const& ts : file_tensors) {
-            if (ts->index_in_zip >= 0) {
+            if (ts.index_in_zip >= 0) {
                 is_zip = true;
                 break;
             }
         }
 
-        std::unique_ptr<MmapWrapper> mmapped;
+        ModelFileData fdata = {};
+        fdata.path          = file_path;
+        fdata.is_zip        = is_zip;
+        fdata.tensors       = std::move(file_tensors);
+
         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());
+            std::unique_ptr<MmapWrapper> mmapped = MmapWrapper::create(file_path, writable_mmap);
+            if (mmapped) {
+                uint8_t* mmap_data             = static_cast<uint8_t*>(mmapped->writable_data());
+                ggml_backend_buffer_t buf_mmap = ggml_backend_cpu_buffer_from_ptr(mmap_data, mmapped->size());
+                if (buf_mmap) {
+                    LOG_INFO("using mmap for '%s'", file_path.c_str());
+                    fdata.mmbuffer = std::shared_ptr<struct ggml_backend_buffer>(buf_mmap, ggml_backend_buffer_free);
+                } else {
+                    LOG_WARN("mmap: failed to create backend buffer for file %s", fdata.path.c_str());
+                }
+                fdata.mmapped = std::shared_ptr<MmapWrapper>(std::move(mmapped));
+            } else {
+                LOG_WARN("failed to memory-map '%s' (falling back to read())", file_path.c_str());
             }
+        } else if (!is_zip) {
+            LOG_INFO("NOT using mmap for '%s' (mmap disabled by caller)",
+                     file_path.c_str());
         }
 
+        file_data.push_back(std::move(fdata));
+    }
+
+    model_files_processed = true;
+
+    int64_t end_time        = ggml_time_ms();
+    int64_t process_time_ms = end_time - start_time;
+
+    LOG_INFO("model files processing completed in %.2fs", process_time_ms / 1000.f);
+}
+
+std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                                       std::set<std::string> ignore_tensors,
+                                                       bool writable_mmap) {
+    process_model_files(true, writable_mmap);
+
+    std::vector<MmapTensorStore> result;
+    uint64_t mapped_bytes = 0;
+    size_t mapped_tensors = 0;
+
+    LOG_DEBUG("memory-mapping tensors...");
+
+    int64_t t_start = ggml_time_ms();
+
+    for (auto& fdata : file_data) {
+        if (!fdata.mmbuffer)
+            continue;
+
+        const std::vector<TensorStorage>& file_tensors = fdata.tensors;
+
+        size_t file_mapped_bytes   = 0;
+        size_t file_mapped_tensors = 0;
+
+        for (const auto& tensor_storage : file_tensors) {
+            const std::string& name = tensor_storage.name;
+
+            bool is_ignored = false;
+            for (const auto& ignore_prefix : ignore_tensors) {
+                if (starts_with(name, ignore_prefix)) {
+                    is_ignored = true;
+                    break;
+                }
+            }
+            if (is_ignored)
+                continue;
+
+            auto it = tensors.find(name);
+            if (it == tensors.end())
+                continue;
+
+            ggml_tensor* dst_tensor = it->second;
+            if (dst_tensor == nullptr)
+                continue;
+
+            if (tensor_storage.type != dst_tensor->type)
+                continue;
+
+            size_t tensor_size   = tensor_storage.nbytes();
+            size_t tensor_offset = tensor_storage.offset;
+
+            if (tensor_storage.ne[0] != dst_tensor->ne[0] ||
+                tensor_storage.ne[1] != dst_tensor->ne[1] ||
+                tensor_storage.ne[2] != dst_tensor->ne[2] ||
+                tensor_storage.ne[3] != dst_tensor->ne[3] ||
+                tensor_size != ggml_nbytes(dst_tensor)) {
+                // let load_tensors worry about this
+                continue;
+            }
+
+            ggml_backend_buffer_t buf_mmap = fdata.mmbuffer.get();
+            uint8_t* mmap_data             = static_cast<uint8_t*>(ggml_backend_buffer_get_base(buf_mmap));
+            dst_tensor->buffer             = buf_mmap;
+            dst_tensor->data               = mmap_data + tensor_offset;
+
+            file_mapped_bytes += tensor_size;
+            file_mapped_tensors++;
+        }
+
+        if (file_mapped_bytes > 0) {
+            mapped_tensors += file_mapped_tensors;
+            mapped_bytes += file_mapped_bytes;
+            result.push_back({fdata.mmapped, fdata.mmbuffer});
+        }
+    }
+
+    int64_t t_end       = ggml_time_ms();
+    int64_t duration_ms = t_end - t_start;
+
+    LOG_INFO("memory-mapped %zu tensors in %zu files (%.2f MB), taking %.2fs",
+             mapped_tensors,
+             result.size(),
+             mapped_bytes / (1024.0 * 1024.0),
+             duration_ms / 1000.0);
+
+    return result;
+}
+
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
+    process_model_files(enable_mmap, false);
+
+    std::atomic<int64_t> read_time_ms(0);
+    std::atomic<int64_t> memcpy_time_ms(0);
+    std::atomic<int64_t> copy_to_backend_time_ms(0);
+    std::atomic<int64_t> convert_time_ms(0);
+    std::atomic<uint64_t> bytes_processed(0);
+
+    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
+    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
+
+    int64_t start_time = ggml_time_ms();
+
+    size_t total_tensors_to_process = 0;
+    for (const auto& fdata : file_data) {
+        total_tensors_to_process += fdata.tensors.size();
+    }
+
+    bool success                   = true;
+    size_t total_tensors_processed = 0;
+    const int64_t t_start          = start_time;
+    int last_n_threads             = 1;
+
+    for (auto& fdata : file_data) {
+        const std::string& file_path = fdata.path;
+        LOG_DEBUG("loading tensors from %s", file_path.c_str());
+
+        const std::vector<TensorStorage>& file_tensors = fdata.tensors;
+
+        bool is_zip = fdata.is_zip;
+
+        std::shared_ptr<MmapWrapper> mmapped = fdata.mmapped;
+
         int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
         if (n_threads < 1) {
             n_threads = 1;
@@ -833,7 +970,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         break;
                     }
 
-                    const TensorStorage& tensor_storage = *file_tensors[idx];
+                    const TensorStorage& tensor_storage = file_tensors[idx];
                     ggml_tensor* dst_tensor             = nullptr;
 
                     t0 = ggml_time_ms();
@@ -850,6 +987,11 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         continue;
                     }
 
+                    // skip mmapped tensors
+                    if (dst_tensor->buffer != nullptr && dst_tensor->buffer == fdata.mmbuffer.get()) {
+                        continue;
+                    }
+
                     size_t nbytes_to_read = tensor_storage.nbytes_to_read();
 
                     auto read_data = [&](char* buf, size_t n) {
@@ -993,9 +1135,8 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     }
 
     int64_t end_time = ggml_time_ms();
-    LOG_INFO("loading tensors completed, taking %.2fs (process: %.2fs, read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
+    LOG_INFO("loading tensors completed, taking %.2fs (read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
              (end_time - start_time) / 1000.f,
-             process_time_ms / 1000.f,
              (read_time_ms.load() / (float)last_n_threads) / 1000.f,
              (memcpy_time_ms.load() / (float)last_n_threads) / 1000.f,
              (convert_time_ms.load() / (float)last_n_threads) / 1000.f,

src/model.h

@@ -43,6 +43,7 @@ enum SDVersion {
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
     VERSION_LTXAV,
+    VERSION_HIDREAM_O1,
     VERSION_Z_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
@@ -172,6 +173,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_sd3(version) ||
         sd_version_is_wan(version) ||
         sd_version_is_qwen_image(version) ||
+        version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
         sd_version_is_ernie_image(version)) {
@@ -202,10 +204,27 @@ using TensorTypeRules = std::vector<std::pair<std::string, ggml_type>>;
 
 TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);
 
+class MmapWrapper;
+
+struct ModelFileData {
+    std::string path;
+    std::vector<TensorStorage> tensors;
+    std::shared_ptr<MmapWrapper> mmapped;
+    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
+    bool is_zip;
+};
+
+struct MmapTensorStore {
+    std::shared_ptr<MmapWrapper> mmapped;
+    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
+};
+
 class ModelLoader {
 protected:
     SDVersion version_ = VERSION_COUNT;
     std::vector<std::string> file_paths_;
+    std::vector<ModelFileData> file_data;
+    bool model_files_processed = false;
     String2TensorStorage tensor_storage_map;
 
     void add_tensor_storage(const TensorStorage& tensor_storage);
@@ -229,6 +248,10 @@ 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 = "");
+    void process_model_files(bool enable_mmap = false, bool writable_mmap = true);
+    std::vector<MmapTensorStore> mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                              std::set<std::string> ignore_tensors = {},
+                                              bool writable                        = true);
     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, ggml_tensor*>& tensors,
                       std::set<std::string> ignore_tensors = {},

src/stable-diffusion.cpp

@@ -1,4 +1,5 @@
 #include "ggml_extend.hpp"
+#include "ggml_graph_cut.h"
 
 #include "model.h"
 #include "rng.hpp"
@@ -55,6 +56,7 @@ const char* model_version_to_str[] = {
     "Flux.2",
     "Flux.2 klein",
     "LTXAV",
+    "HiDream O1",
     "Z-Image",
     "Ovis Image",
     "Ernie Image",
@@ -76,6 +78,8 @@ const char* sampling_methods_str[] = {
     "Res Multistep",
     "Res 2s",
     "ER-SDE",
+    "Euler CFG++",
+    "Euler A CFG++",
 };
 
 /*================================================== Helper Functions ================================================*/
@@ -111,6 +115,7 @@ static float get_cache_reuse_threshold(const sd_cache_params_t& params) {
 
 class StableDiffusionGGML {
 public:
+    std::vector<MmapTensorStore> mmap_tensor_store;
     ggml_backend_t backend             = nullptr;  // general backend
     ggml_backend_t clip_backend        = nullptr;
     ggml_backend_t control_net_backend = nullptr;
@@ -214,6 +219,7 @@ public:
         ggml_log_set(ggml_log_callback_default, nullptr);
 
         init_backend();
+        max_vram = sd::ggml_graph_cut::resolve_max_vram_gib(max_vram, backend);
 
         ModelLoader model_loader;
 
@@ -385,6 +391,51 @@ public:
             apply_lora_immediately = false;
         }
 
+        std::map<std::string, ggml_tensor*> mmap_able_tensors;
+        bool enable_mmap_tensors = false;
+        bool main_backend_mmap   = false;
+        bool needs_writable_mmap = false;
+        if (sd_ctx_params->enable_mmap) {
+            if (apply_lora_immediately) {
+                needs_writable_mmap = true;
+                LOG_WARN("in mode 'immediately', LoRAs will cause extra memory usage with mmap");
+            }
+            enable_mmap_tensors = true;
+            if (offload_params_to_cpu) {
+                main_backend_mmap = true;
+            } else {
+                ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+                struct ggml_backend_dev_props props;
+                ggml_backend_dev_get_props(dev, &props);
+                main_backend_mmap = props.caps.buffer_from_host_ptr;
+            }
+        }
+
+        // split definition to avoid msvc choking on the extra parameter handling
+        auto get_param_tensors_p = [&](auto&& model, bool force_cpu, const char* prefix) {
+            std::map<std::string, ggml_tensor*> temp;
+            model->get_param_tensors(temp, prefix);
+            bool do_mmap = enable_mmap_tensors && (main_backend_mmap || force_cpu);
+            for (const auto& [key, tensor] : temp) {
+                tensors[key] = tensor;
+                if (do_mmap) {
+                    mmap_able_tensors[key] = tensor;
+                }
+            }
+        };
+
+        auto get_param_tensors = [&](auto&& model, bool force_cpu = false) {
+            std::map<std::string, ggml_tensor*> temp;
+            model->get_param_tensors(temp);
+            bool do_mmap = enable_mmap_tensors && (main_backend_mmap || force_cpu);
+            for (const auto& [key, tensor] : temp) {
+                tensors[key] = tensor;
+                if (do_mmap) {
+                    mmap_able_tensors[key] = tensor;
+                }
+            }
+        };
+
         if (sd_version_is_control(version)) {
             // Might need vae encode for control cond
             vae_decode_only = false;
@@ -405,9 +456,7 @@ public:
 
         bool clip_on_cpu = sd_ctx_params->keep_clip_on_cpu;
 
-        const size_t max_graph_vram_bytes = max_vram <= 0.f
-                                                ? 0
-                                                : static_cast<size_t>(static_cast<double>(max_vram) * 1024.0 * 1024.0 * 1024.0);
+        const size_t max_graph_vram_bytes = sd::ggml_graph_cut::max_vram_gib_to_bytes(max_vram);
 
         {
             clip_backend = backend;
@@ -507,8 +556,7 @@ public:
                                                                              offload_params_to_cpu,
                                                                              tensor_storage_map);
                     clip_vision->set_max_graph_vram_bytes(max_graph_vram_bytes);
-                    clip_vision->alloc_params_buffer();
-                    clip_vision->get_param_tensors(tensors);
+                    get_param_tensors(clip_vision);
                 }
             } else if (sd_version_is_qwen_image(version)) {
                 bool enable_vision = false;
@@ -527,6 +575,14 @@ public:
                                                                    "model.diffusion_model",
                                                                    version,
                                                                    sd_ctx_params->qwen_image_zero_cond_t);
+            } else if (version == VERSION_HIDREAM_O1) {
+                cond_stage_model = std::make_shared<HiDreamO1::HiDreamO1Conditioner>(clip_backend,
+                                                                                     offload_params_to_cpu,
+                                                                                     tensor_storage_map);
+                diffusion_model  = std::make_shared<HiDreamO1Model>(backend,
+                                                                   offload_params_to_cpu,
+                                                                   tensor_storage_map,
+                                                                   "model");
             } else if (sd_version_is_anima(version)) {
                 cond_stage_model = std::make_shared<AnimaConditioner>(clip_backend,
                                                                       offload_params_to_cpu,
@@ -584,12 +640,10 @@ public:
             }
 
             cond_stage_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
-            cond_stage_model->alloc_params_buffer();
-            cond_stage_model->get_param_tensors(tensors);
+            get_param_tensors(cond_stage_model, clip_on_cpu);
 
             diffusion_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
-            diffusion_model->alloc_params_buffer();
-            diffusion_model->get_param_tensors(tensors);
+            get_param_tensors(diffusion_model);
 
             if (sd_version_is_unet_edit(version)) {
                 vae_decode_only = false;
@@ -597,8 +651,7 @@ public:
 
             if (high_noise_diffusion_model) {
                 high_noise_diffusion_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
-                high_noise_diffusion_model->alloc_params_buffer();
-                high_noise_diffusion_model->get_param_tensors(tensors);
+                get_param_tensors(high_noise_diffusion_model);
             }
 
             if (sd_ctx_params->keep_vae_on_cpu && !ggml_backend_is_cpu(backend)) {
@@ -668,7 +721,9 @@ public:
                 }
             };
 
-            if (version == VERSION_CHROMA_RADIANCE) {
+            bool force_vae_cpu = sd_ctx_params->keep_vae_on_cpu;
+
+            if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1) {
                 LOG_INFO("using FakeVAE");
                 first_stage_model = std::make_shared<FakeVAE>(version,
                                                               vae_backend,
@@ -677,20 +732,17 @@ public:
                 LOG_INFO("using TAE for encoding / decoding");
                 first_stage_model = create_tae();
                 first_stage_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
-                first_stage_model->alloc_params_buffer();
-                first_stage_model->get_param_tensors(tensors, "tae");
+                get_param_tensors_p(first_stage_model, force_vae_cpu, "tae");
             } else {
                 LOG_INFO("using VAE for encoding / decoding");
                 first_stage_model = create_vae();
                 first_stage_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
-                first_stage_model->alloc_params_buffer();
-                first_stage_model->get_param_tensors(tensors, "first_stage_model");
+                get_param_tensors_p(first_stage_model, force_vae_cpu, "first_stage_model");
                 if (use_tae && tae_preview_only) {
                     LOG_INFO("using TAE for preview");
                     preview_vae = create_tae();
                     preview_vae->set_max_graph_vram_bytes(max_graph_vram_bytes);
-                    preview_vae->alloc_params_buffer();
-                    preview_vae->get_param_tensors(tensors, "tae");
+                    get_param_tensors_p(first_stage_model, force_vae_cpu, "vae");
                 }
             }
 
@@ -769,11 +821,7 @@ public:
                 }
             }
             if (use_pmid) {
-                if (!pmid_model->alloc_params_buffer()) {
-                    LOG_ERROR(" pmid model params buffer allocation failed");
-                    return false;
-                }
-                pmid_model->get_param_tensors(tensors, "pmid");
+                get_param_tensors_p(pmid_model, false, "pmid");
             }
 
             if (sd_ctx_params->flash_attn) {
@@ -856,6 +904,45 @@ public:
             ignore_tensors.insert("text_encoders.llm.vision_tower.");
             ignore_tensors.insert("text_encoders.llm.multi_modal_projector.");
         }
+        if (version == VERSION_HIDREAM_O1) {
+            ignore_tensors.insert("lm_head.");
+            ignore_tensors.insert("model.visual.deepstack_merger_list.");
+        }
+
+        if (enable_mmap_tensors) {
+            if (mmap_able_tensors.empty()) {
+                LOG_DEBUG("no tensors could be memory-mapped");
+            } else {
+                mmap_tensor_store = model_loader.mmap_tensors(mmap_able_tensors, ignore_tensors, needs_writable_mmap);
+            }
+        }
+
+        if (clip_vision) {
+            clip_vision->alloc_params_buffer();
+        }
+        if (cond_stage_model) {
+            cond_stage_model->alloc_params_buffer();
+        }
+        if (diffusion_model) {
+            diffusion_model->alloc_params_buffer();
+        }
+        if (high_noise_diffusion_model) {
+            high_noise_diffusion_model->alloc_params_buffer();
+        }
+        if (first_stage_model) {
+            first_stage_model->alloc_params_buffer();
+        }
+        if (preview_vae) {
+            preview_vae->alloc_params_buffer();
+        }
+        if (use_pmid && pmid_model) {
+            if (!pmid_model->alloc_params_buffer()) {
+                LOG_ERROR(" pmid model params buffer allocation failed");
+                ggml_free(ctx);
+                return false;
+            }
+        }
+
         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");
@@ -962,6 +1049,7 @@ public:
                            sd_version_is_ltxav(version) ||
                            sd_version_is_wan(version) ||
                            sd_version_is_qwen_image(version) ||
+                           version == VERSION_HIDREAM_O1 ||
                            sd_version_is_anima(version) ||
                            sd_version_is_ernie_image(version) ||
                            sd_version_is_z_image(version)) {
@@ -1568,6 +1656,9 @@ public:
         if (sd_version_is_anima(version)) {
             return std::vector<float>{t / static_cast<float>(TIMESTEPS)};
         }
+        if (version == VERSION_HIDREAM_O1) {
+            return std::vector<float>{1.0f - (t / static_cast<float>(TIMESTEPS))};
+        }
         if (sd_version_is_z_image(version)) {
             return std::vector<float>{1000.f - t};
         }
@@ -1656,6 +1747,7 @@ public:
                              int shifted_timestep,
                              sample_method_t method,
                              bool is_flow_denoiser,
+                             const char* extra_sample_args,
                              const std::vector<float>& sigmas,
                              int start_merge_step,
                              const std::vector<sd::Tensor<float>>& ref_latents,
@@ -1674,13 +1766,26 @@ public:
                                                                                            cache_params,
                                                                                            denoiser.get(),
                                                                                            sigmas);
-        size_t steps                                = sigmas.size() - 1;
-        bool has_skiplayer                          = slg_scale != 0.0f && !skip_layers.empty();
+
+        // Spectrum cache is not supported for CFG++ samplers
+        if (method == EULER_CFG_PP_SAMPLE_METHOD || method == EULER_A_CFG_PP_SAMPLE_METHOD) {
+            if (cache_runtime.spectrum_enabled) {
+                LOG_WARN("Spectrum cache requested but not supported for CFG++ samplers");
+                cache_runtime.spectrum_enabled = false;
+            }
+        }
+
+        size_t steps       = sigmas.size() - 1;
+        bool has_skiplayer = slg_scale != 0.0f && !skip_layers.empty();
         if (has_skiplayer && !sd_version_is_dit(version)) {
             has_skiplayer = false;
             LOG_WARN("SLG is incompatible with this model type");
         }
 
+        if (version == VERSION_HIDREAM_O1 && !noise.empty()) {
+            noise *= eta;
+        }
+
         int64_t t0                   = ggml_time_us();
         sd::Tensor<float> x_t        = !noise.empty()
                                            ? denoiser->noise_scaling(sigmas[0], noise, init_latent)
@@ -1688,7 +1793,7 @@ public:
         sd::Tensor<float> denoised   = x_t;
         SamplePreviewContext preview = prepare_sample_preview_context();
 
-        auto denoise = [&](const sd::Tensor<float>& x, float sigma, int step) -> sd::Tensor<float> {
+        auto denoise = [&](const sd::Tensor<float>& x, float sigma, int step, sd::Tensor<float>* out_uncond_denoised = nullptr) -> sd::Tensor<float> {
             if (step == 1 || step == -1) {
                 pretty_progress(0, (int)steps, 0);
             }
@@ -1711,15 +1816,17 @@ public:
             }
 
             if (cache_runtime.spectrum_enabled && cache_runtime.spectrum.should_predict()) {
-                cache_runtime.spectrum.predict(&denoised);
-                if (!denoise_mask.empty()) {
-                    denoised = denoised * denoise_mask + init_latent * (1.0f - denoise_mask);
+                if (out_uncond_denoised == nullptr) {
+                    cache_runtime.spectrum.predict(&denoised);
+                    if (!denoise_mask.empty()) {
+                        denoised = denoised * denoise_mask + init_latent * (1.0f - denoise_mask);
+                    }
+                    if (sd_should_preview_denoised() && preview.callback != nullptr) {
+                        preview_image(step, denoised, version, preview.mode, preview.callback, preview.data, false);
+                    }
+                    report_sample_progress(step, steps, t0);
+                    return denoised;
                 }
-                if (sd_should_preview_denoised() && preview.callback != nullptr) {
-                    preview_image(step, denoised, version, preview.mode, preview.callback, preview.data, false);
-                }
-                report_sample_progress(step, steps, t0);
-                return denoised;
             }
 
             if (sd_should_preview_noisy() && preview.callback != nullptr) {
@@ -1754,12 +1861,18 @@ public:
             auto run_condition = [&](const SDCondition& condition,
                                      const sd::Tensor<float>* c_concat_override = nullptr,
                                      const std::vector<int>* local_skip_layers  = nullptr) -> sd::Tensor<float> {
-                diffusion_params.context     = condition.c_crossattn.empty() ? nullptr : &condition.c_crossattn;
-                diffusion_params.c_concat    = c_concat_override != nullptr ? c_concat_override : (condition.c_concat.empty() ? nullptr : &condition.c_concat);
-                diffusion_params.y           = condition.c_vector.empty() ? nullptr : &condition.c_vector;
-                diffusion_params.t5_ids      = condition.c_t5_ids.empty() ? nullptr : &condition.c_t5_ids;
-                diffusion_params.t5_weights  = condition.c_t5_weights.empty() ? nullptr : &condition.c_t5_weights;
-                diffusion_params.skip_layers = local_skip_layers;
+                diffusion_params.context      = condition.c_crossattn.empty() ? nullptr : &condition.c_crossattn;
+                diffusion_params.c_concat     = c_concat_override != nullptr ? c_concat_override : (condition.c_concat.empty() ? nullptr : &condition.c_concat);
+                diffusion_params.y            = condition.c_vector.empty() ? nullptr : &condition.c_vector;
+                diffusion_params.t5_ids       = condition.c_t5_ids.empty() ? nullptr : &condition.c_t5_ids;
+                diffusion_params.t5_weights   = condition.c_t5_weights.empty() ? nullptr : &condition.c_t5_weights;
+                diffusion_params.input_ids    = condition.c_input_ids.empty() ? nullptr : &condition.c_input_ids;
+                diffusion_params.input_pos    = condition.c_position_ids.empty() ? nullptr : &condition.c_position_ids;
+                diffusion_params.token_types  = condition.c_token_types.empty() ? nullptr : &condition.c_token_types;
+                diffusion_params.vinput_mask  = condition.c_vinput_mask.empty() ? nullptr : &condition.c_vinput_mask;
+                diffusion_params.image_embeds = condition.c_image_embeds.empty() ? nullptr : &condition.c_image_embeds;
+                diffusion_params.ref_latents  = condition.c_ref_images.empty() ? &ref_latents : &condition.c_ref_images;
+                diffusion_params.skip_layers  = local_skip_layers;
 
                 sd::Tensor<float> cached_output;
                 if (step_cache.before_condition(&condition, noised_input, &cached_output)) {
@@ -1843,6 +1956,10 @@ public:
                 latent_result += (cond_out - skip_cond_out) * slg_scale;
             }
             denoised = latent_result * c_out + x * c_skip;
+            if (out_uncond_denoised != nullptr) {
+                sd::Tensor<float> base_uncond = !uncond_out.empty() ? uncond_out : cond_out;
+                *out_uncond_denoised          = base_uncond * c_out + x * c_skip;
+            }
             if (cache_runtime.spectrum_enabled) {
                 cache_runtime.spectrum.update(denoised);
             }
@@ -1856,7 +1973,7 @@ public:
             return denoised;
         };
 
-        auto x0_opt = sample_k_diffusion(method, denoise, x_t, sigmas, sampler_rng, eta, is_flow_denoiser);
+        auto x0_opt = sample_k_diffusion(method, denoise, x_t, sigmas, sampler_rng, eta, is_flow_denoiser, extra_sample_args);
         if (x0_opt.empty()) {
             LOG_ERROR("Diffusion model sampling failed");
             if (control_net) {
@@ -1908,6 +2025,8 @@ public:
                 latent_channel = 128;
             } else if (version == VERSION_WAN2_2_TI2V) {
                 latent_channel = 48;
+            } else if (version == VERSION_HIDREAM_O1) {
+                latent_channel = 3;
             } else if (version == VERSION_CHROMA_RADIANCE) {
                 latent_channel = 3;
             } else if (sd_version_uses_flux2_vae(version)) {
@@ -2056,6 +2175,8 @@ const char* sample_method_to_str[] = {
     "res_multistep",
     "res_2s",
     "er_sde",
+    "euler_cfg_pp",
+    "euler_a_cfg_pp",
 };
 
 const char* sd_sample_method_name(enum sample_method_t sample_method) {
@@ -2365,6 +2486,7 @@ void sd_sample_params_init(sd_sample_params_t* sample_params) {
     sample_params->custom_sigmas               = nullptr;
     sample_params->custom_sigmas_count         = 0;
     sample_params->flow_shift                  = INFINITY;
+    sample_params->extra_sample_args           = nullptr;
 }
 
 char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
@@ -2386,7 +2508,8 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
              "sample_steps: %d, "
              "eta: %.2f, "
              "shifted_timestep: %d, "
-             "flow_shift: %.2f)",
+             "flow_shift: %.2f, "
+             "extra_sample_args: %s)",
              sample_params->guidance.txt_cfg,
              std::isfinite(sample_params->guidance.img_cfg)
                  ? sample_params->guidance.img_cfg
@@ -2401,7 +2524,8 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
              sample_params->sample_steps,
              sample_params->eta,
              sample_params->shifted_timestep,
-             sample_params->flow_shift);
+             sample_params->flow_shift,
+             SAFE_STR(sample_params->extra_sample_args));
 
     return buf;
 }
@@ -2654,6 +2778,9 @@ static float resolve_eta(sd_ctx_t* sd_ctx,
                          float eta,
                          enum sample_method_t sample_method) {
     if (eta == INFINITY) {
+        if (sd_ctx->sd->version == VERSION_HIDREAM_O1) {
+            return 8.f;
+        }
         switch (sample_method) {
             case DDIM_TRAILING_SAMPLE_METHOD:
             case TCD_SAMPLE_METHOD:
@@ -2663,6 +2790,7 @@ static float resolve_eta(sd_ctx_t* sd_ctx,
             case EULER_A_SAMPLE_METHOD:
             case DPMPP2S_A_SAMPLE_METHOD:
             case ER_SDE_SAMPLE_METHOD:
+            case EULER_A_CFG_PP_SAMPLE_METHOD:
                 return 1.0f;
             default:;
         }
@@ -2886,6 +3014,8 @@ struct GenerationRequest {
 struct SamplePlan {
     enum sample_method_t sample_method            = SAMPLE_METHOD_COUNT;
     enum sample_method_t high_noise_sample_method = SAMPLE_METHOD_COUNT;
+    const char* extra_sample_args                 = nullptr;
+    const char* high_noise_extra_sample_args      = nullptr;
     float eta                                     = 0.f;
     float high_noise_eta                          = 0.f;
     int sample_steps                              = 0;
@@ -2898,22 +3028,25 @@ struct SamplePlan {
     SamplePlan(sd_ctx_t* sd_ctx,
                const sd_img_gen_params_t* sd_img_gen_params,
                const GenerationRequest& request) {
-        sample_method = sd_img_gen_params->sample_params.sample_method;
-        eta           = sd_img_gen_params->sample_params.eta;
-        sample_steps  = sd_img_gen_params->sample_params.sample_steps;
+        sample_method     = sd_img_gen_params->sample_params.sample_method;
+        extra_sample_args = sd_img_gen_params->sample_params.extra_sample_args;
+        eta               = sd_img_gen_params->sample_params.eta;
+        sample_steps      = sd_img_gen_params->sample_params.sample_steps;
         resolve(sd_ctx, &request, &sd_img_gen_params->sample_params);
     }
 
     SamplePlan(sd_ctx_t* sd_ctx,
                const sd_vid_gen_params_t* sd_vid_gen_params,
                const GenerationRequest& request) {
-        sample_method = sd_vid_gen_params->sample_params.sample_method;
-        eta           = sd_vid_gen_params->sample_params.eta;
-        sample_steps  = sd_vid_gen_params->sample_params.sample_steps;
+        sample_method     = sd_vid_gen_params->sample_params.sample_method;
+        extra_sample_args = sd_vid_gen_params->sample_params.extra_sample_args;
+        eta               = sd_vid_gen_params->sample_params.eta;
+        sample_steps      = sd_vid_gen_params->sample_params.sample_steps;
         if (sd_ctx->sd->high_noise_diffusion_model) {
-            high_noise_sample_steps  = sd_vid_gen_params->high_noise_sample_params.sample_steps;
-            high_noise_sample_method = sd_vid_gen_params->high_noise_sample_params.sample_method;
-            high_noise_eta           = sd_vid_gen_params->high_noise_sample_params.eta;
+            high_noise_sample_steps      = sd_vid_gen_params->high_noise_sample_params.sample_steps;
+            high_noise_sample_method     = sd_vid_gen_params->high_noise_sample_params.sample_method;
+            high_noise_extra_sample_args = sd_vid_gen_params->high_noise_sample_params.extra_sample_args;
+            high_noise_eta               = sd_vid_gen_params->high_noise_sample_params.eta;
         }
         moe_boundary = sd_vid_gen_params->moe_boundary;
         resolve(sd_ctx, &request, &sd_vid_gen_params->sample_params);
@@ -3234,6 +3367,9 @@ static std::optional<ImageGenerationLatents> prepare_image_generation_latents(sd
 
     std::vector<sd::Tensor<float>> ref_latents;
     for (size_t i = 0; i < ref_images.size(); i++) {
+        if (sd_ctx->sd->version == VERSION_HIDREAM_O1) {
+            continue;
+        }
         sd::Tensor<float> ref_latent;
         if (request->auto_resize_ref_image) {
             LOG_DEBUG("auto resize ref images");
@@ -3644,6 +3780,7 @@ SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* s
                                                    request.shifted_timestep,
                                                    plan.sample_method,
                                                    sd_ctx->sd->is_flow_denoiser(),
+                                                   plan.extra_sample_args,
                                                    plan.sigmas,
                                                    plan.start_merge_step,
                                                    latents.ref_latents,
@@ -3686,9 +3823,7 @@ SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* s
             hires_upscaler                    = std::make_unique<UpscalerGGML>(sd_ctx->sd->n_threads,
                                                             false,
                                                             request.hires.upscale_tile_size);
-            const size_t max_graph_vram_bytes = sd_ctx->sd->max_vram <= 0.f
-                                                    ? 0
-                                                    : static_cast<size_t>(static_cast<double>(sd_ctx->sd->max_vram) * 1024.0 * 1024.0 * 1024.0);
+            const size_t max_graph_vram_bytes = sd::ggml_graph_cut::max_vram_gib_to_bytes(sd_ctx->sd->max_vram);
             hires_upscaler->set_max_graph_vram_bytes(max_graph_vram_bytes);
             if (!hires_upscaler->load_from_file(request.hires.model_path,
                                                 sd_ctx->sd->offload_params_to_cpu,
@@ -3770,6 +3905,7 @@ SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* s
                                                             request.shifted_timestep,
                                                             plan.sample_method,
                                                             sd_ctx->sd->is_flow_denoiser(),
+                                                            plan.extra_sample_args,
                                                             hires_sigma_sched,
                                                             plan.start_merge_step,
                                                             latents.ref_latents,
@@ -4184,6 +4320,7 @@ SD_API bool generate_video(sd_ctx_t* sd_ctx,
                                                            request.shifted_timestep,
                                                            plan.high_noise_sample_method,
                                                            sd_ctx->sd->is_flow_denoiser(),
+                                                           plan.high_noise_extra_sample_args,
                                                            high_noise_sigmas,
                                                            -1,
                                                            std::vector<sd::Tensor<float>>{},
@@ -4227,6 +4364,7 @@ SD_API bool generate_video(sd_ctx_t* sd_ctx,
                                                         sd_vid_gen_params->sample_params.shifted_timestep,
                                                         plan.sample_method,
                                                         sd_ctx->sd->is_flow_denoiser(),
+                                                        plan.extra_sample_args,
                                                         plan.sigmas,
                                                         -1,
                                                         std::vector<sd::Tensor<float>>{},

src/tokenizers/qwen2_tokenizer.cpp

@@ -81,6 +81,11 @@ Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str) {
         "</tool_response>",
         "<think>",
         "</think>",
+        "<|boi_token|>",
+        "<|bor_token|>",
+        "<|eor_token|>",
+        "<|bot_token|>",
+        "<|tms_token|>",
     };
 
     if (merges_utf8_str.size() > 0) {

src/util.cpp

@@ -112,7 +112,7 @@ private:
     HANDLE hmapping_;
 };
 
-std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bool writable) {
     void* mapped_data = nullptr;
     size_t file_size  = 0;
 
@@ -137,14 +137,18 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
 
     file_size = static_cast<size_t>(size.QuadPart);
 
-    HANDLE mapping_handle = CreateFileMapping(file_handle, nullptr, PAGE_READONLY, 0, 0, nullptr);
+    DWORD page_prot = writable ? PAGE_WRITECOPY : PAGE_READONLY;
+
+    HANDLE mapping_handle = CreateFileMapping(file_handle, nullptr, page_prot, 0, 0, nullptr);
 
     if (mapping_handle == nullptr) {
         CloseHandle(file_handle);
         return nullptr;
     }
 
-    mapped_data = MapViewOfFile(mapping_handle, FILE_MAP_READ, 0, 0, file_size);
+    DWORD view_access = writable ? FILE_MAP_COPY : FILE_MAP_READ;
+
+    mapped_data = MapViewOfFile(mapping_handle, view_access, 0, 0, file_size);
 
     if (mapped_data == nullptr) {
         CloseHandle(mapping_handle);
@@ -172,28 +176,85 @@ 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_);
-    }
+struct MmapFlags {
+    bool sequential;
+    bool populate;
+    bool willneed;
+    bool dontneed;
 };
 
-std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
+static MmapFlags get_mmap_flags() {
+    MmapFlags result          = {};
+    const char* SD_MMAP_FLAGS = std::getenv("SD_MMAP_FLAGS");
+    if (SD_MMAP_FLAGS && *SD_MMAP_FLAGS) {
+        std::stringstream ss(SD_MMAP_FLAGS);
+        std::string token;
+        while (std::getline(ss, token, ',')) {
+            std::string ntoken = trim(token);
+            std::transform(ntoken.begin(), ntoken.end(), ntoken.begin(), ::tolower);
+            if (ntoken == "sequential") {
+                result.sequential = true;
+            } else if (ntoken == "populate") {
+                result.populate = true;
+            } else if (ntoken == "willneed") {
+                result.willneed = true;
+            } else if (ntoken == "dontneed") {
+                result.dontneed = true;
+            }
+        }
+    }
+    return result;
+}
+
+class MmapWrapperImpl : public MmapWrapper {
+public:
+    MmapWrapperImpl(void* data, size_t size, int fd)
+        : MmapWrapper(data, size), fd_(fd) {}
+
+    ~MmapWrapperImpl() override {
+#ifdef __linux__
+        auto cfg_flags = get_mmap_flags();
+
+        // Drop the kernel pagecache pages for this file. madvise(DONTNEED)
+        // alone only unmaps from the process address space; pagecache
+        // entries persist (`free` reports them as buff/cache and the OOM
+        // killer doesn't touch them, but they ARE counted against
+        // overcommit and can starve other allocations on tight-RAM
+        // systems). posix_fadvise(POSIX_FADV_DONTNEED) is the documented
+        // way to evict pagecache for a specific fd's pages.
+        if (cfg_flags.dontneed) {
+            madvise(data_, size_, MADV_DONTNEED);
+            posix_fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED);
+        }
+#endif
+        munmap(data_, size_);
+        close(fd_);
+    }
+
+private:
+    int fd_;
+};
+
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bool writable) {
     int file_descriptor = open(filename.c_str(), O_RDONLY);
     if (file_descriptor == -1) {
         return nullptr;
     }
 
+    auto cfg_flags = get_mmap_flags();
+
     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;
+    // Sequential access hint helps the kernel read-ahead efficiently and
+    // also encourages eviction of already-read pages (the kernel keeps
+    // a smaller working set when this is set).
+    if (cfg_flags.sequential) {
+        posix_fadvise(file_descriptor, 0, 0, POSIX_FADV_SEQUENTIAL);
+    }
+    if (cfg_flags.populate) {
+        mmap_flags |= MAP_POPULATE;
+    }
 #endif
 
     struct stat sb;
@@ -204,20 +265,27 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
 
     size_t file_size = sb.st_size;
 
-    void* mapped_data = mmap(nullptr, file_size, PROT_READ, mmap_flags, file_descriptor, 0);
+    if (file_size == 0) {
+        close(file_descriptor);
+        return nullptr;
+    }
 
-    close(file_descriptor);
+    int mmap_prot = PROT_READ | (writable ? PROT_WRITE : 0);
+
+    void* mapped_data = mmap(nullptr, file_size, mmap_prot, mmap_flags, file_descriptor, 0);
 
     if (mapped_data == MAP_FAILED) {
+        close(file_descriptor);
         return nullptr;
     }
 
 #ifdef __linux__
-    // performance flags used by llama.cpp
-    // posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
+    if (cfg_flags.willneed) {
+        posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
+    }
 #endif
 
-    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size);
+    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size, file_descriptor);
 }
 
 #endif

src/util.h

@@ -42,7 +42,7 @@ sd::Tensor<float> clip_preprocess(const sd::Tensor<float>& image, int target_wid
 
 class MmapWrapper {
 public:
-    static std::unique_ptr<MmapWrapper> create(const std::string& filename);
+    static std::unique_ptr<MmapWrapper> create(const std::string& filename, bool writable = false);
 
     virtual ~MmapWrapper() = default;
 
@@ -52,6 +52,7 @@ public:
     MmapWrapper& operator=(MmapWrapper&&)      = delete;
 
     const uint8_t* data() const { return static_cast<uint8_t*>(data_); }
+    uint8_t* writable_data() { return static_cast<uint8_t*>(data_); }
     size_t size() const { return size_; }
     bool copy_data(void* buf, size_t n, size_t offset) const;
 

src/vae.hpp

@@ -73,7 +73,7 @@ public:
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
-        } else if (version == VERSION_CHROMA_RADIANCE) {
+        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1) {
             scale_factor = 1;
         }
         return scale_factor;