feat: add max-vram based segmented param offload (#1476)

Co-authored-by: Stéphane du Hamel <stephduh@live.fr>
Co-authored-by: Cyberhan123 <255542417@qq.com>
Co-authored-by: leejet <leejet714@gmail.com>

CMakeLists.txt

@@ -72,37 +72,31 @@ option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF
 if(SD_CUDA)
     message("-- Use CUDA as backend stable-diffusion")
     set(GGML_CUDA ON)
-    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_METAL)
     message("-- Use Metal as backend stable-diffusion")
     set(GGML_METAL ON)
-    add_definitions(-DSD_USE_METAL)
 endif()
 
 if (SD_VULKAN)
     message("-- Use Vulkan as backend stable-diffusion")
     set(GGML_VULKAN ON)
-    add_definitions(-DSD_USE_VULKAN)
 endif ()
 
 if (SD_OPENCL)
     message("-- Use OpenCL as backend stable-diffusion")
     set(GGML_OPENCL ON)
-    add_definitions(-DSD_USE_OPENCL)
 endif ()
 
 if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
-    add_definitions(-DSD_USE_CUDA)
 endif ()
 
 if(SD_MUSA)
     message("-- Use MUSA as backend stable-diffusion")
     set(GGML_MUSA ON)
-    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_WEBP)
@@ -222,7 +216,6 @@ if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-    add_definitions(-DSD_USE_SYCL)
     # disable fast-math on host, see:
     # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
     if (WIN32)

docs/caching.md

@@ -131,8 +131,6 @@ sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
 | `warmup` | Steps to always compute before caching starts | 4 |
 | `stop` | Stop caching at this fraction of total steps | 0.9 |
 
-```
-
 ### Performance Tips
 
 - Start with default thresholds and adjust based on output quality

examples/cli/README.md

@@ -4,19 +4,19 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
-                              ./output.png) (eg. output_%03d.png). For video generation, single-file outputs support .avi, .webm, and animated .webp
-  --preview-path <string>     path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp
-  --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
-                              every step)
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
+  -o, --output <string>         path to write result image to. you can use printf-style %d format specifiers for image
+                                sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs
+                                support .avi, .webm, and animated .webp
   --image <string>              path to the image to inspect (for metadata mode)
   --metadata-format <string>    metadata output format, one of [text, json] (default: text)
+  --preview-path <string>       path to write preview image to (default: ./preview.png). Multi-frame previews support
+                                .avi, .webm, and animated .webp
+  --preview-interval <int>      interval in denoising steps between consecutive updates of the image preview file
+                                (default is 1, meaning updating at every step)
+  --output-begin-idx <int>      starting index for output image sequence, must be non-negative (default 0 if specified
+                                %d in output path, 1 otherwise)
   --canny                       apply canny preprocessor (edge detection)
   --convert-name                convert tensor name (for convert mode)
-                              convert mode writes `.gguf` or `.safetensors` based on the output extension.
-                              `.safetensors` export currently supports f16, bf16, f32, and i32 tensor types only.
-                              i32 is passthrough only; no f32 <-> i32 conversion is performed
   -v, --verbose                 print extra info
   --color                       colors the logging tags according to level
   --taesd-preview-only          prevents usage of taesd for decoding the final image. (for use with --preview tae)
@@ -34,7 +34,8 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -46,16 +47,18 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
+  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           CPU physical cores
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+                                           then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --vae-tiling                             process vae in tiles to reduce memory usage
+  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
+                                           graph splitting
   --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
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -70,20 +73,19 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           type of the weight file
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+                                           q4_K). If not specified, the default is the type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           immediately will be used.The immediately mode may have precision and
-                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           other hand, is exactly the opposite.
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
-                                           (overrides --vae-tile-size)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           immediately mode may have precision and compatibility issues with quantized
+                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+                                           opposite.
 
 Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -92,69 +94,99 @@ Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           lexicographical (character) order. For example, if the control video path is
-                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+                                           inside should be stored as images in lexicographical (character) order. For
+                                           example, if the control video path is `frames`, the directory contain images
+                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --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)
   -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)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           will be 1 for SD1.x, 2 for SD2.x
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Vibrant
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
+  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
+  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
+  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
+                                           128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
+                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           medium
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+                                           disabled, a value of 2.5 is nice for sd3.5 medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
+                                           res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
-  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+                                           (default: 3.5)
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
+                                           0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
+                                           res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>              
-  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
+                                           destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
+                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
+  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
+  --increase-ref-index                     automatically increase the indices of references images based on the order
+                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --hires                                  enable highres fix
   -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)
-  --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
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm, bong_tangent], default: discrete
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --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)
+  --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
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
+                                           discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
+                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level),
-                                           'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=;
-                                           spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0" or "w=0.4,window=2"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
+                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```
 
 Metadata mode inspects PNG/JPEG container metadata without loading any model:

examples/cli/main.cpp

@@ -278,7 +278,9 @@ void parse_args(int argc, const char** argv, SDCliParams& cli_params, SDContextP
     bool valid = cli_params.resolve_and_validate();
     if (valid && cli_params.mode != METADATA) {
         valid = ctx_params.resolve_and_validate(cli_params.mode) &&
-                gen_params.resolve_and_validate(cli_params.mode, ctx_params.lora_model_dir);
+                gen_params.resolve_and_validate(cli_params.mode,
+                                                ctx_params.lora_model_dir,
+                                                ctx_params.hires_upscalers_dir);
     }
 
     if (!valid) {
@@ -431,8 +433,9 @@ bool save_results(const SDCliParams& cli_params,
         if (!img.data)
             return false;
 
+        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx;
         std::string params          = gen_params.embed_image_metadata
-                                 ? get_image_params(ctx_params, gen_params, gen_params.seed + idx)
+                                          ? get_image_params(ctx_params, gen_params, metadata_seed, cli_params.mode)
                                           : "";
         const bool ok               = write_image_to_file(path.string(), img.data, img.width, img.height, img.channel, params, 90);
         LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
@@ -688,6 +691,13 @@ int main(int argc, const char* argv[]) {
         vae_decode_only = false;
     }
 
+    if (gen_params.hires_enabled &&
+        (gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_MODEL ||
+         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_LANCZOS ||
+         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_NEAREST)) {
+        vae_decode_only = false;
+    }
+
     sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, true, cli_params.taesd_preview);
 
     SDImageVec results;

examples/common/common.cpp

@@ -107,47 +107,60 @@ static bool is_absolute_path(const std::string& p) {
 
 std::string ArgOptions::wrap_text(const std::string& text, size_t width, size_t indent) {
     std::ostringstream oss;
-    size_t line_len = 0;
     size_t pos      = 0;
+    size_t line_len = 0;
 
     while (pos < text.size()) {
-        // Preserve manual newlines
         if (text[pos] == '\n') {
             oss << '\n'
                 << std::string(indent, ' ');
-            line_len = indent;
+            line_len = 0;
             ++pos;
             continue;
         }
 
-        // Add the character
-        oss << text[pos];
-        ++line_len;
+        if (std::isspace(static_cast<unsigned char>(text[pos]))) {
             ++pos;
-
-        // If the current line exceeds width, try to break at the last space
-        if (line_len >= width) {
-            std::string current = oss.str();
-            size_t back         = current.size();
-
-            // Find the last space (for a clean break)
-            while (back > 0 && current[back - 1] != ' ' && current[back - 1] != '\n')
-                --back;
-
-            // If found a space to break on
-            if (back > 0 && current[back - 1] != '\n') {
-                std::string before = current.substr(0, back - 1);
-                std::string after  = current.substr(back);
-                oss.str("");
-                oss.clear();
-                oss << before << "\n"
-                    << std::string(indent, ' ') << after;
-            } else {
-                // If no space found, just break at width
-                oss << "\n"
-                    << std::string(indent, ' ');
+            continue;
+        }
+
+        size_t word_start = pos;
+        while (pos < text.size() &&
+               text[pos] != '\n' &&
+               !std::isspace(static_cast<unsigned char>(text[pos]))) {
+            ++pos;
+        }
+
+        std::string word = text.substr(word_start, pos - word_start);
+        while (!word.empty()) {
+            size_t separator_len = line_len == 0 ? 0 : 1;
+            if (line_len + separator_len + word.size() <= width) {
+                if (separator_len > 0) {
+                    oss << ' ';
+                    ++line_len;
+                }
+                oss << word;
+                line_len += word.size();
+                word.clear();
+                continue;
+            }
+
+            if (line_len > 0) {
+                oss << '\n'
+                    << std::string(indent, ' ');
+                line_len = 0;
+                continue;
+            }
+
+            size_t chunk_len = std::min(width, word.size());
+            oss << word.substr(0, chunk_len);
+            line_len = chunk_len;
+            word.erase(0, chunk_len);
+            if (!word.empty()) {
+                oss << '\n'
+                    << std::string(indent, ' ');
+                line_len = 0;
             }
-            line_len = indent;
         }
     }
 
@@ -351,7 +364,10 @@ ArgOptions SDContextParams::get_options() {
          "--lora-model-dir",
          "lora model directory",
          &lora_model_dir},
-
+        {"",
+         "--hires-upscalers-dir",
+         "highres fix upscaler model directory",
+         &hires_upscalers_dir},
         {"",
          "--tensor-type-rules",
          "weight type per tensor pattern (example: \"^vae\\.=f16,model\\.=q8_0\")",
@@ -378,7 +394,12 @@ ArgOptions SDContextParams::get_options() {
          &chroma_t5_mask_pad},
     };
 
-    options.float_options = {};
+    options.float_options = {
+        {"",
+         "--max-vram",
+         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting",
+         &max_vram},
+    };
 
     options.bool_options = {
         {"",
@@ -649,10 +670,12 @@ std::string SDContextParams::to_string() const {
         << "  wtype: " << sd_type_name(wtype) << ",\n"
         << "  tensor_type_rules: \"" << tensor_type_rules << "\",\n"
         << "  lora_model_dir: \"" << lora_model_dir << "\",\n"
+        << "  hires_upscalers_dir: \"" << hires_upscalers_dir << "\",\n"
         << "  photo_maker_path: \"" << photo_maker_path << "\",\n"
         << "  rng_type: " << sd_rng_type_name(rng_type) << ",\n"
         << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
         << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
+        << "  max_vram: " << max_vram << ",\n"
         << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
         << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
         << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
@@ -727,6 +750,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
         chroma_use_t5_mask,
         chroma_t5_mask_pad,
         qwen_image_zero_cond_t,
+        max_vram,
     };
     return sd_ctx_params;
 }
@@ -777,6 +801,12 @@ ArgOptions SDGenerationParams::get_options() {
          "--pm-id-embed-path",
          "path to PHOTOMAKER v2 id embed",
          &pm_id_embed_path},
+        {"",
+         "--hires-upscaler",
+         "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)",
+         &hires_upscaler},
     };
 
     options.int_options = {
@@ -826,6 +856,22 @@ ArgOptions SDGenerationParams::get_options() {
          "--upscale-tile-size",
          "tile size for ESRGAN upscaling (default: 128)",
          &upscale_tile_size},
+        {"",
+         "--hires-width",
+         "highres fix target width, 0 to use --hires-scale (default: 0)",
+         &hires_width},
+        {"",
+         "--hires-height",
+         "highres fix target height, 0 to use --hires-scale (default: 0)",
+         &hires_height},
+        {"",
+         "--hires-steps",
+         "highres fix second pass sample steps, 0 to reuse --steps (default: 0)",
+         &hires_steps},
+        {"",
+         "--hires-upscale-tile-size",
+         "highres fix upscaler tile size, reserved for model-backed upscalers (default: 128)",
+         &hires_upscale_tile_size},
     };
 
     options.float_options = {
@@ -913,6 +959,14 @@ ArgOptions SDGenerationParams::get_options() {
          "--vae-tile-overlap",
          "tile overlap for vae tiling, in fraction of tile size (default: 0.5)",
          &vae_tiling_params.target_overlap},
+        {"",
+         "--hires-scale",
+         "highres fix scale when target size is not set (default: 2.0)",
+         &hires_scale},
+        {"",
+         "--hires-denoising-strength",
+         "highres fix second pass denoising strength (default: 0.7)",
+         &hires_denoising_strength},
     };
 
     options.bool_options = {
@@ -936,6 +990,11 @@ ArgOptions SDGenerationParams::get_options() {
          "process vae in tiles to reduce memory usage",
          true,
          &vae_tiling_params.enabled},
+        {"",
+         "--hires",
+         "enable highres fix",
+         true,
+         &hires_enabled},
     };
 
     auto on_seed_arg = [&](int argc, const char** argv, int index) {
@@ -1424,6 +1483,37 @@ static bool parse_lora_json_field(const json& parent,
     return true;
 }
 
+static bool resolve_model_file_from_dir(const std::string& model_name,
+                                        const std::string& model_dir,
+                                        const std::vector<std::string>& valid_ext,
+                                        const char* label,
+                                        std::string& resolved_path) {
+    if (model_dir.empty()) {
+        LOG_ERROR("%s directory is empty", label);
+        return false;
+    }
+    if (model_name.empty() ||
+        model_name.find('/') != std::string::npos ||
+        model_name.find('\\') != std::string::npos ||
+        fs::path(model_name).has_root_path() ||
+        fs::path(model_name).has_extension()) {
+        LOG_ERROR("%s must be a model name without path or extension: %s", label, model_name.c_str());
+        return false;
+    }
+
+    fs::path model_dir_path = model_dir;
+    for (const auto& ext : valid_ext) {
+        fs::path try_path = model_dir_path / (model_name + ext);
+        if (fs::exists(try_path) && fs::is_regular_file(try_path)) {
+            resolved_path = try_path.lexically_normal().string();
+            return true;
+        }
+    }
+
+    LOG_ERROR("can not find %s %s in %s", label, model_name.c_str(), model_dir_path.lexically_normal().string().c_str());
+    return false;
+}
+
 bool SDGenerationParams::from_json_str(
     const std::string& json_str,
     const std::function<std::string(const std::string&)>& lora_path_resolver) {
@@ -1487,6 +1577,34 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("increase_ref_index", increase_ref_index);
     load_if_exists("embed_image_metadata", embed_image_metadata);
 
+    if (j.contains("hires") && j["hires"].is_object()) {
+        const json& hires_json = j["hires"];
+        if (hires_json.contains("enabled") && hires_json["enabled"].is_boolean()) {
+            hires_enabled = hires_json["enabled"];
+        }
+        if (hires_json.contains("upscaler") && hires_json["upscaler"].is_string()) {
+            hires_upscaler = hires_json["upscaler"];
+        }
+        if (hires_json.contains("scale") && hires_json["scale"].is_number()) {
+            hires_scale = hires_json["scale"];
+        }
+        if (hires_json.contains("target_width") && hires_json["target_width"].is_number_integer()) {
+            hires_width = hires_json["target_width"];
+        }
+        if (hires_json.contains("target_height") && hires_json["target_height"].is_number_integer()) {
+            hires_height = hires_json["target_height"];
+        }
+        if (hires_json.contains("steps") && hires_json["steps"].is_number_integer()) {
+            hires_steps = hires_json["steps"];
+        }
+        if (hires_json.contains("denoising_strength") && hires_json["denoising_strength"].is_number()) {
+            hires_denoising_strength = hires_json["denoising_strength"];
+        }
+        if (hires_json.contains("upscale_tile_size") && hires_json["upscale_tile_size"].is_number_integer()) {
+            hires_upscale_tile_size = hires_json["upscale_tile_size"];
+        }
+    }
+
     auto parse_sample_params_json = [&](const json& sample_json,
                                         sd_sample_params_t& target_params,
                                         std::vector<int>& target_skip_layers,
@@ -1800,7 +1918,7 @@ bool SDGenerationParams::initialize_cache_params() {
     return true;
 }
 
-bool SDGenerationParams::resolve(const std::string& lora_model_dir, bool strict) {
+bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict) {
     if (high_noise_sample_params.sample_steps <= 0) {
         high_noise_sample_params.sample_steps = -1;
     }
@@ -1819,6 +1937,27 @@ bool SDGenerationParams::resolve(const std::string& lora_model_dir, bool strict)
         sample_params.sample_steps = std::clamp(sample_params.sample_steps, 1, 100);
     }
 
+    hires_upscaler_model_path.clear();
+    if (hires_enabled) {
+        if (hires_upscaler.empty()) {
+            hires_upscaler = "Latent";
+        }
+        resolved_hires_upscaler = str_to_sd_hires_upscaler(hires_upscaler.c_str());
+        if (resolved_hires_upscaler == SD_HIRES_UPSCALER_NONE) {
+            hires_enabled = false;
+        } else if (resolved_hires_upscaler == SD_HIRES_UPSCALER_COUNT) {
+            static const std::vector<std::string> valid_ext = {".gguf", ".safetensors", ".pt", ".pth"};
+            if (!resolve_model_file_from_dir(hires_upscaler,
+                                             hires_upscalers_dir,
+                                             valid_ext,
+                                             "hires upscaler",
+                                             hires_upscaler_model_path)) {
+                return false;
+            }
+            resolved_hires_upscaler = SD_HIRES_UPSCALER_MODEL;
+        }
+    }
+
     prompt_with_lora = prompt;
     if (!lora_model_dir.empty()) {
         extract_and_remove_lora(lora_model_dir);
@@ -1883,6 +2022,29 @@ bool SDGenerationParams::validate(SDMode mode) {
         return false;
     }
 
+    if (hires_enabled) {
+        if (hires_width < 0 || hires_height < 0) {
+            LOG_ERROR("error: hires target width and height must be >= 0");
+            return false;
+        }
+        if (hires_scale <= 0.f && hires_width <= 0 && hires_height <= 0) {
+            LOG_ERROR("error: hires scale must be positive when target size is not set");
+            return false;
+        }
+        if (hires_steps < 0) {
+            LOG_ERROR("error: hires steps must be >= 0");
+            return false;
+        }
+        if (hires_denoising_strength <= 0.f || hires_denoising_strength > 1.f) {
+            LOG_ERROR("error: hires denoising strength must be in (0.0, 1.0]");
+            return false;
+        }
+        if (hires_upscale_tile_size < 1) {
+            LOG_ERROR("error: hires upscale tile size must be positive");
+            return false;
+        }
+    }
+
     if (mode == UPSCALE) {
         if (init_image_path.length() == 0) {
             LOG_ERROR("error: upscale mode needs an init image (--init-img)\n");
@@ -1893,8 +2055,11 @@ bool SDGenerationParams::validate(SDMode mode) {
     return true;
 }
 
-bool SDGenerationParams::resolve_and_validate(SDMode mode, const std::string& lora_model_dir, bool strict) {
-    if (!resolve(lora_model_dir, strict)) {
+bool SDGenerationParams::resolve_and_validate(SDMode mode,
+                                              const std::string& lora_model_dir,
+                                              const std::string& hires_upscalers_dir,
+                                              bool strict) {
+    if (!resolve(lora_model_dir, hires_upscalers_dir, strict)) {
         return false;
     }
     if (!validate(mode)) {
@@ -1965,6 +2130,16 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.pm_params             = pm_params;
     params.vae_tiling_params     = vae_tiling_params;
     params.cache                 = cache_params;
+
+    params.hires.enabled            = hires_enabled;
+    params.hires.upscaler           = resolved_hires_upscaler;
+    params.hires.model_path         = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
+    params.hires.scale              = hires_scale;
+    params.hires.target_width       = hires_width;
+    params.hires.target_height      = hires_height;
+    params.hires.steps              = hires_steps;
+    params.hires.denoising_strength = hires_denoising_strength;
+    params.hires.upscale_tile_size  = hires_upscale_tile_size;
     return params;
 }
 
@@ -2089,6 +2264,15 @@ std::string SDGenerationParams::to_string() const {
         << "  seed: " << seed << ",\n"
         << "  upscale_repeats: " << upscale_repeats << ",\n"
         << "  upscale_tile_size: " << upscale_tile_size << ",\n"
+        << "  hires: { enabled: " << (hires_enabled ? "true" : "false")
+        << ", upscaler: \"" << hires_upscaler << "\""
+        << ", model_path: \"" << hires_upscaler_model_path << "\""
+        << ", scale: " << hires_scale
+        << ", target_width: " << hires_width
+        << ", target_height: " << hires_height
+        << ", steps: " << hires_steps
+        << ", denoising_strength: " << hires_denoising_strength
+        << ", upscale_tile_size: " << hires_upscale_tile_size << " },\n"
         << "  vae_tiling_params: { "
         << vae_tiling_params.enabled << ", "
         << vae_tiling_params.tile_size_x << ", "
@@ -2104,7 +2288,192 @@ std::string version_string() {
     return std::string("stable-diffusion.cpp version ") + sd_version() + ", commit " + sd_commit();
 }
 
-std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed) {
+static std::string safe_json_string(const char* value) {
+    return value ? value : "";
+}
+
+static void set_json_basename_if_not_empty(json& target, const char* key, const std::string& path) {
+    if (!path.empty()) {
+        target[key] = sd_basename(path);
+    }
+}
+
+static json build_sampling_metadata_json(const sd_sample_params_t& sample_params,
+                                         const std::vector<int>& skip_layers,
+                                         const std::vector<float>* custom_sigmas = nullptr) {
+    json sampling = {
+        {"steps", sample_params.sample_steps},
+        {"eta", sample_params.eta},
+        {"shifted_timestep", sample_params.shifted_timestep},
+        {"flow_shift", sample_params.flow_shift},
+        {"guidance",
+         {
+             {"txt_cfg", sample_params.guidance.txt_cfg},
+             {"img_cfg", sample_params.guidance.img_cfg},
+             {"distilled_guidance", sample_params.guidance.distilled_guidance},
+             {"slg",
+              {
+                  {"scale", sample_params.guidance.slg.scale},
+                  {"layers", skip_layers},
+                  {"start", sample_params.guidance.slg.layer_start},
+                  {"end", sample_params.guidance.slg.layer_end},
+              }},
+         }},
+    };
+    if (sample_params.sample_method != SAMPLE_METHOD_COUNT) {
+        sampling["method"] = safe_json_string(sd_sample_method_name(sample_params.sample_method));
+    }
+    if (sample_params.scheduler != SCHEDULER_COUNT) {
+        sampling["scheduler"] = safe_json_string(sd_scheduler_name(sample_params.scheduler));
+    }
+    if (custom_sigmas != nullptr) {
+        sampling["custom_sigmas"] = *custom_sigmas;
+    }
+    return sampling;
+}
+
+std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
+                                            const SDGenerationParams& gen_params,
+                                            int64_t seed,
+                                            SDMode mode) {
+    json root;
+    root["schema"]    = "sdcpp.image.params/v1";
+    root["mode"]      = mode == VID_GEN ? "vid_gen" : "img_gen";
+    root["generator"] = {
+        {"name", "stable-diffusion.cpp"},
+        {"version", safe_json_string(sd_version())},
+        {"commit", safe_json_string(sd_commit())},
+    };
+    root["seed"]   = seed;
+    root["width"]  = gen_params.get_resolved_width();
+    root["height"] = gen_params.get_resolved_height();
+
+    root["prompt"] = {
+        {"positive", gen_params.prompt},
+        {"negative", gen_params.negative_prompt},
+    };
+    root["sampling"] = build_sampling_metadata_json(gen_params.sample_params,
+                                                    gen_params.skip_layers,
+                                                    &gen_params.custom_sigmas);
+
+    json models;
+    set_json_basename_if_not_empty(models, "model", ctx_params.model_path);
+    set_json_basename_if_not_empty(models, "clip_l", ctx_params.clip_l_path);
+    set_json_basename_if_not_empty(models, "clip_g", ctx_params.clip_g_path);
+    set_json_basename_if_not_empty(models, "clip_vision", ctx_params.clip_vision_path);
+    set_json_basename_if_not_empty(models, "t5xxl", ctx_params.t5xxl_path);
+    set_json_basename_if_not_empty(models, "llm", ctx_params.llm_path);
+    set_json_basename_if_not_empty(models, "llm_vision", ctx_params.llm_vision_path);
+    set_json_basename_if_not_empty(models, "diffusion_model", ctx_params.diffusion_model_path);
+    set_json_basename_if_not_empty(models, "high_noise_diffusion_model", ctx_params.high_noise_diffusion_model_path);
+    set_json_basename_if_not_empty(models, "vae", ctx_params.vae_path);
+    set_json_basename_if_not_empty(models, "taesd", ctx_params.taesd_path);
+    set_json_basename_if_not_empty(models, "control_net", ctx_params.control_net_path);
+    root["models"] = std::move(models);
+
+    root["clip_skip"]             = gen_params.clip_skip;
+    root["strength"]              = gen_params.strength;
+    root["control_strength"]      = gen_params.control_strength;
+    root["auto_resize_ref_image"] = gen_params.auto_resize_ref_image;
+    root["increase_ref_index"]    = gen_params.increase_ref_index;
+    if (mode == VID_GEN) {
+        root["video"] = {
+            {"frame_count", gen_params.video_frames},
+            {"fps", gen_params.fps},
+        };
+        root["moe_boundary"]        = gen_params.moe_boundary;
+        root["vace_strength"]       = gen_params.vace_strength;
+        root["high_noise_sampling"] = build_sampling_metadata_json(gen_params.high_noise_sample_params,
+                                                                   gen_params.high_noise_skip_layers);
+    }
+
+    root["rng"] = safe_json_string(sd_rng_type_name(ctx_params.rng_type));
+    if (ctx_params.sampler_rng_type != RNG_TYPE_COUNT) {
+        root["sampler_rng"] = safe_json_string(sd_rng_type_name(ctx_params.sampler_rng_type));
+    }
+
+    json loras = json::array();
+    for (const auto& entry : gen_params.lora_map) {
+        loras.push_back({
+            {"name", sd_basename(entry.first)},
+            {"multiplier", entry.second},
+            {"is_high_noise", false},
+        });
+    }
+    for (const auto& entry : gen_params.high_noise_lora_map) {
+        loras.push_back({
+            {"name", sd_basename(entry.first)},
+            {"multiplier", entry.second},
+            {"is_high_noise", true},
+        });
+    }
+    if (!loras.empty()) {
+        root["loras"] = std::move(loras);
+    }
+
+    if (gen_params.hires_enabled) {
+        root["hires"] = {
+            {"enabled", gen_params.hires_enabled},
+            {"upscaler", gen_params.hires_upscaler},
+            {"model", gen_params.hires_upscaler_model_path.empty() ? "" : sd_basename(gen_params.hires_upscaler_model_path)},
+            {"scale", gen_params.hires_scale},
+            {"target_width", gen_params.hires_width},
+            {"target_height", gen_params.hires_height},
+            {"steps", gen_params.hires_steps},
+            {"denoising_strength", gen_params.hires_denoising_strength},
+            {"upscale_tile_size", gen_params.hires_upscale_tile_size},
+        };
+    }
+
+    if (gen_params.cache_params.mode != SD_CACHE_DISABLED) {
+        root["cache"] = {
+            {"requested_mode", gen_params.cache_mode},
+            {"requested_option", gen_params.cache_option},
+            {"mode", gen_params.cache_params.mode},
+            {"scm_mask", gen_params.scm_mask},
+            {"scm_policy_dynamic", gen_params.scm_policy_dynamic},
+            {"reuse_threshold", gen_params.cache_params.reuse_threshold},
+            {"start_percent", gen_params.cache_params.start_percent},
+            {"end_percent", gen_params.cache_params.end_percent},
+            {"error_decay_rate", gen_params.cache_params.error_decay_rate},
+            {"use_relative_threshold", gen_params.cache_params.use_relative_threshold},
+            {"reset_error_on_compute", gen_params.cache_params.reset_error_on_compute},
+            {"Fn_compute_blocks", gen_params.cache_params.Fn_compute_blocks},
+            {"Bn_compute_blocks", gen_params.cache_params.Bn_compute_blocks},
+            {"residual_diff_threshold", gen_params.cache_params.residual_diff_threshold},
+            {"max_warmup_steps", gen_params.cache_params.max_warmup_steps},
+            {"max_cached_steps", gen_params.cache_params.max_cached_steps},
+            {"max_continuous_cached_steps", gen_params.cache_params.max_continuous_cached_steps},
+            {"taylorseer_n_derivatives", gen_params.cache_params.taylorseer_n_derivatives},
+            {"taylorseer_skip_interval", gen_params.cache_params.taylorseer_skip_interval},
+            {"spectrum_w", gen_params.cache_params.spectrum_w},
+            {"spectrum_m", gen_params.cache_params.spectrum_m},
+            {"spectrum_lam", gen_params.cache_params.spectrum_lam},
+            {"spectrum_window_size", gen_params.cache_params.spectrum_window_size},
+            {"spectrum_flex_window", gen_params.cache_params.spectrum_flex_window},
+            {"spectrum_warmup_steps", gen_params.cache_params.spectrum_warmup_steps},
+            {"spectrum_stop_percent", gen_params.cache_params.spectrum_stop_percent},
+        };
+    }
+
+    if (gen_params.vae_tiling_params.enabled) {
+        root["vae_tiling"] = {
+            {"enabled", gen_params.vae_tiling_params.enabled},
+            {"tile_size_x", gen_params.vae_tiling_params.tile_size_x},
+            {"tile_size_y", gen_params.vae_tiling_params.tile_size_y},
+            {"target_overlap", gen_params.vae_tiling_params.target_overlap},
+            {"rel_size_x", gen_params.vae_tiling_params.rel_size_x},
+            {"rel_size_y", gen_params.vae_tiling_params.rel_size_y},
+        };
+    }
+
+    return root.dump();
+}
+
+std::string get_image_params(const SDContextParams& ctx_params,
+                             const SDGenerationParams& gen_params,
+                             int64_t seed,
+                             SDMode mode) {
     std::string parameter_string;
     if (gen_params.prompt_with_lora.size() != 0) {
         parameter_string += gen_params.prompt_with_lora + "\n";
@@ -2117,7 +2486,7 @@ std::string get_image_params(const SDContextParams& ctx_params, const SDGenerati
     parameter_string += "Steps: " + std::to_string(gen_params.sample_params.sample_steps) + ", ";
     parameter_string += "CFG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
     if (gen_params.sample_params.guidance.slg.scale != 0 && gen_params.skip_layers.size() != 0) {
-        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
+        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.slg.scale) + ", ";
         parameter_string += "Skip layers: [";
         for (const auto& layer : gen_params.skip_layers) {
             parameter_string += std::to_string(layer) + ", ";
@@ -2162,6 +2531,14 @@ std::string get_image_params(const SDContextParams& ctx_params, const SDGenerati
     if (gen_params.clip_skip != -1) {
         parameter_string += "Clip skip: " + std::to_string(gen_params.clip_skip) + ", ";
     }
+    if (gen_params.hires_enabled) {
+        parameter_string += "Hires upscale: " + gen_params.hires_upscaler + ", ";
+        parameter_string += "Hires scale: " + std::to_string(gen_params.hires_scale) + ", ";
+        parameter_string += "Hires resize: " + std::to_string(gen_params.hires_width) + "x" + std::to_string(gen_params.hires_height) + ", ";
+        parameter_string += "Hires steps: " + std::to_string(gen_params.hires_steps) + ", ";
+        parameter_string += "Denoising strength: " + std::to_string(gen_params.hires_denoising_strength) + ", ";
+    }
     parameter_string += "Version: stable-diffusion.cpp";
+    parameter_string += ", SDCPP: " + build_sdcpp_image_metadata_json(ctx_params, gen_params, seed, mode);
     return parameter_string;
 }

examples/common/common.h

@@ -101,6 +101,7 @@ struct SDContextParams {
     sd_type_t wtype = SD_TYPE_COUNT;
     std::string tensor_type_rules;
     std::string lora_model_dir = ".";
+    std::string hires_upscalers_dir;
 
     std::map<std::string, std::string> embedding_map;
     std::vector<sd_embedding_t> embedding_vec;
@@ -108,6 +109,7 @@ struct SDContextParams {
     rng_type_t rng_type         = CUDA_RNG;
     rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
     bool offload_params_to_cpu  = false;
+    float max_vram              = 0.f;
     bool enable_mmap            = false;
     bool control_net_cpu        = false;
     bool clip_on_cpu            = false;
@@ -190,12 +192,23 @@ struct SDGenerationParams {
     int upscale_repeats   = 1;
     int upscale_tile_size = 128;
 
+    bool hires_enabled         = false;
+    std::string hires_upscaler = "Latent";
+    std::string hires_upscaler_model_path;
+    float hires_scale              = 2.f;
+    int hires_width                = 0;
+    int hires_height               = 0;
+    int hires_steps                = 0;
+    float hires_denoising_strength = 0.7f;
+    int hires_upscale_tile_size    = 128;
+
     std::map<std::string, float> lora_map;
     std::map<std::string, float> high_noise_lora_map;
 
     // Derived and normalized fields.
     std::string prompt_with_lora;  // for metadata record only
     std::vector<sd_lora_t> lora_vec;
+    sd_hires_upscaler_t resolved_hires_upscaler;
 
     // Owned execution payload.
     SDImageOwner init_image;
@@ -225,15 +238,25 @@ struct SDGenerationParams {
     void set_width_and_height_if_unset(int w, int h);
     int get_resolved_width() const;
     int get_resolved_height() const;
-    bool resolve(const std::string& lora_model_dir, bool strict = false);
+    bool resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict = false);
     bool validate(SDMode mode);
-    bool resolve_and_validate(SDMode mode, const std::string& lora_model_dir, bool strict = false);
+    bool resolve_and_validate(SDMode mode,
+                              const std::string& lora_model_dir,
+                              const std::string& hires_upscalers_dir,
+                              bool strict = false);
     sd_img_gen_params_t to_sd_img_gen_params_t();
     sd_vid_gen_params_t to_sd_vid_gen_params_t();
     std::string to_string() const;
 };
 
 std::string version_string();
-std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed);
+std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
+                                            const SDGenerationParams& gen_params,
+                                            int64_t seed,
+                                            SDMode mode = IMG_GEN);
+std::string get_image_params(const SDContextParams& ctx_params,
+                             const SDGenerationParams& gen_params,
+                             int64_t seed,
+                             SDMode mode = IMG_GEN);
 
 #endif  // __EXAMPLES_COMMON_COMMON_H__

examples/server/README.md

@@ -136,7 +136,8 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -148,16 +149,18 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
+  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           CPU physical cores
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+                                           then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --vae-tiling                             process vae in tiles to reduce memory usage
+  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
+                                           graph splitting
   --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
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -172,20 +175,19 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           type of the weight file
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+                                           q4_K). If not specified, the default is the type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           immediately will be used.The immediately mode may have precision and
-                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           other hand, is exactly the opposite.
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
-                                           (overrides --vae-tile-size)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           immediately mode may have precision and compatibility issues with quantized
+                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+                                           opposite.
 
 Default Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -194,65 +196,97 @@ Default Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           lexicographical (character) order. For example, if the control video path is
-                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+                                           inside should be stored as images in lexicographical (character) order. For
+                                           example, if the control video path is `frames`, the directory contain images
+                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --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)
   -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)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           will be 1 for SD1.x, 2 for SD2.x
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Vibrant
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
+  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
+  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
+  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
+                                           128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
+                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           medium
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+                                           disabled, a value of 2.5 is nice for sd3.5 medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
+                                           res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
-  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+                                           (default: 3.5)
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
+                                           0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
+                                           res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>              
-  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
+                                           destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
+                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
+  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
+  --increase-ref-index                     automatically increase the indices of references images based on the order
+                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --hires                                  enable highres fix
   -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)
-  --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
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm, bong_tangent], default: discrete
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --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)
+  --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
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
+                                           discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
+                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
+                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```

examples/server/api.md

@@ -38,6 +38,8 @@ Current generation-related endpoints include:
 - `POST /sdapi/v1/txt2img`
 - `POST /sdapi/v1/img2img`
 - `GET /sdapi/v1/loras`
+- `GET /sdapi/v1/upscalers`
+- `GET /sdapi/v1/latent-upscale-modes`
 - `GET /sdapi/v1/samplers`
 - `GET /sdapi/v1/schedulers`
 - `GET /sdapi/v1/sd-models`
@@ -216,6 +218,13 @@ Currently supported request fields:
 | `scheduler` | `string` | Scheduler name |
 | `lora` | `array<object>` | Structured LoRA list |
 | `extra_images` | `array<string>` | Base64 or data URL images |
+| `enable_hr` | `boolean` | Enable highres fix for `txt2img` |
+| `hr_upscaler` | `string` | `Lanczos`, `Nearest`, a latent mode such as `Latent (nearest-exact)`, or an upscaler model name from `/sdapi/v1/upscalers` |
+| `hr_scale` | `number` | Highres scale when resize target is not set |
+| `hr_resize_x` | `integer` | Highres target width, `0` to use scale |
+| `hr_resize_y` | `integer` | Highres target height, `0` to use scale |
+| `hr_steps` | `integer` | Highres second-pass sample steps, `0` to reuse `steps` |
+| `denoising_strength` | `number` | Highres denoising strength for `txt2img` |
 
 Native extension fields:
 
@@ -241,6 +250,8 @@ Currently supported request fields:
 | `inpainting_mask_invert` | `integer` or `boolean` | Treated as invert flag |
 | `denoising_strength` | `number` | Clamped to `0.0..1.0` |
 
+Highres fix fields are currently handled for `txt2img`; `img2img` uses `denoising_strength` as image-to-image strength.
+
 Native extension fields:
 
 - any `sdcpp API` fields embedded through `sd_cpp_extra_args` inside `prompt`
@@ -258,6 +269,8 @@ Response fields:
 Currently exposed:
 
 - `GET /sdapi/v1/loras`
+- `GET /sdapi/v1/upscalers`
+- `GET /sdapi/v1/latent-upscale-modes`
 - `GET /sdapi/v1/samplers`
 - `GET /sdapi/v1/schedulers`
 - `GET /sdapi/v1/sd-models`
@@ -272,6 +285,26 @@ Response fields:
 | `[].name` | `string` | Display name derived from file stem |
 | `[].path` | `string` | Relative path under the configured LoRA directory |
 
+`GET /sdapi/v1/upscalers`
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `[].name` | `string` | Built-in name or model stem |
+| `[].model_name` | `string \| null` | Model family label for model-backed upscalers |
+| `[].model_path` | `string \| null` | Absolute model path for model-backed upscalers |
+| `[].model_url` | `string \| null` | Currently always null |
+| `[].scale` | `integer` | Currently `4` |
+
+Built-in entries include `None`, `Lanczos`, and `Nearest`. Model-backed entries are scanned from the top level of `--hires-upscalers-dir`; subdirectories are not scanned.
+
+`GET /sdapi/v1/latent-upscale-modes`
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `[].name` | `string` | WebUI-compatible latent upscale mode name |
+
+Built-in latent modes include `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, and `Latent (bicubic antialiased)`.
+
 `GET /sdapi/v1/samplers`
 
 | Field | Type | Notes |
@@ -388,6 +421,7 @@ Top-level fields:
 | `samplers` | `array<string>` | Available sampling methods |
 | `schedulers` | `array<string>` | Available schedulers |
 | `loras` | `array<object>` | Available LoRA entries |
+| `upscalers` | `array<object>` | Available model-backed highres upscalers |
 | `limits` | `object` | Shared queue and size limits |
 
 `model`
@@ -424,6 +458,14 @@ Shared nested fields:
 | `loras[].name` | `string` |
 | `loras[].path` | `string` |
 
+`upscalers`
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `upscalers[].name` | `string` | Built-in name or model stem; use this value in `hires.upscaler` |
+
+Built-in entries include `None`, `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, and `Latent (bicubic antialiased)`. Model-backed entries are scanned from the top level of `--hires-upscalers-dir`; subdirectories are not scanned.
+
 `limits`
 
 | Field | Type |
@@ -482,6 +524,15 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `auto_resize_ref_image` | `boolean` |
 | `increase_ref_index` | `boolean` |
 | `control_strength` | `number` |
+| `hires` | `object` |
+| `hires.enabled` | `boolean` |
+| `hires.upscaler` | `string` |
+| `hires.scale` | `number` |
+| `hires.target_width` | `integer` |
+| `hires.target_height` | `integer` |
+| `hires.steps` | `integer` |
+| `hires.denoising_strength` | `number` |
+| `hires.upscale_tile_size` | `integer` |
 
 `vid_gen`-specific default fields:
 
@@ -514,6 +565,7 @@ Fields returned in `features_by_mode.img_gen`:
 - `ref_images`
 - `lora`
 - `vae_tiling`
+- `hires`
 - `cache`
 - `cancel_queued`
 - `cancel_generating`
@@ -625,6 +677,16 @@ Example:
   },
 
   "lora": [],
+  "hires": {
+    "enabled": false,
+    "upscaler": "Latent",
+    "scale": 2.0,
+    "target_width": 0,
+    "target_height": 0,
+    "steps": 0,
+    "denoising_strength": 0.7,
+    "upscale_tile_size": 128
+  },
 
   "vae_tiling_params": {
     "enabled": false,
@@ -729,12 +791,23 @@ Other native fields:
 
 | Field | Type |
 | --- | --- |
+| `hires` | `object` |
+| `hires.enabled` | `boolean` |
+| `hires.upscaler` | `string` |
+| `hires.scale` | `number` |
+| `hires.target_width` | `integer` |
+| `hires.target_height` | `integer` |
+| `hires.steps` | `integer` |
+| `hires.denoising_strength` | `number` |
+| `hires.upscale_tile_size` | `integer` |
 | `vae_tiling_params` | `object` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
 | `scm_policy_dynamic` | `boolean` |
 
+For `hires.upscaler`, use `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, `Latent (bicubic antialiased)`, or an `upscalers[].name` value from `GET /sdcpp/v1/capabilities`. Model-backed upscalers are resolved as `--hires-upscalers-dir / (name + ext)` and must live directly in that directory.
+
 HTTP-only output fields:
 
 | Field | Type |

examples/server/main.cpp

@@ -48,7 +48,9 @@ static void parse_args(int argc,
 
     if (!svr_params.resolve_and_validate() ||
         !ctx_params.resolve_and_validate(IMG_GEN) ||
-        !default_gen_params.resolve_and_validate(IMG_GEN, ctx_params.lora_model_dir)) {
+        !default_gen_params.resolve_and_validate(IMG_GEN,
+                                                 ctx_params.lora_model_dir,
+                                                 ctx_params.hires_upscalers_dir)) {
         print_usage(argv[0], options_vec);
         exit(1);
     }
@@ -95,6 +97,8 @@ int main(int argc, const char** argv) {
 
     std::vector<LoraEntry> lora_cache;
     std::mutex lora_mutex;
+    std::vector<UpscalerEntry> upscaler_cache;
+    std::mutex upscaler_mutex;
     AsyncJobManager async_job_manager;
     ServerRuntime runtime = {
         sd_ctx.get(),
@@ -104,6 +108,8 @@ int main(int argc, const char** argv) {
         &default_gen_params,
         &lora_cache,
         &lora_mutex,
+        &upscaler_cache,
+        &upscaler_mutex,
         &async_job_manager,
     };
 

examples/server/routes_openai.cpp

@@ -70,7 +70,7 @@ static bool build_openai_generation_request(const httplib::Request& req,
     }
 
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid params";
         return false;
     }
@@ -212,7 +212,7 @@ static bool build_openai_edit_request(const httplib::Request& req,
     }
 
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid params";
         return false;
     }

examples/server/routes_sdapi.cpp

@@ -1,6 +1,7 @@
 #include "routes.h"
 
 #include <algorithm>
+#include <cctype>
 #include <cstring>
 #include <regex>
 #include <string_view>
@@ -35,14 +36,20 @@ static fs::path resolve_display_model_path(const ServerRuntime& runtime) {
     return {};
 }
 
+static std::string lower_ascii(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
 static enum sample_method_t get_sdapi_sample_method(std::string name) {
     enum sample_method_t result = str_to_sample_method(name.c_str());
     if (result != SAMPLE_METHOD_COUNT) {
         return result;
     }
 
-    std::transform(name.begin(), name.end(), name.begin(),
-                   [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
+    name = lower_ascii(name);
     static const std::unordered_map<std::string_view, sample_method_t> hardcoded{
         {"euler a", EULER_A_SAMPLE_METHOD},
         {"k_euler_a", EULER_A_SAMPLE_METHOD},
@@ -114,6 +121,18 @@ static bool build_sdapi_img_gen_request(const json& j,
     request.gen_params.width                          = j.value("width", -1);
     request.gen_params.height                         = j.value("height", -1);
 
+    if (!img2img && j.value("enable_hr", false)) {
+        request.gen_params.hires_enabled = true;
+        request.gen_params.hires_scale   = j.value("hr_scale", request.gen_params.hires_scale);
+        request.gen_params.hires_width   = j.value("hr_resize_x", request.gen_params.hires_width);
+        request.gen_params.hires_height  = j.value("hr_resize_y", request.gen_params.hires_height);
+        request.gen_params.hires_steps   = j.value("hr_steps", request.gen_params.hires_steps);
+        request.gen_params.hires_denoising_strength =
+            j.value("denoising_strength", request.gen_params.hires_denoising_strength);
+
+        request.gen_params.hires_upscaler = j.value("hr_upscaler", request.gen_params.hires_upscaler);
+    }
+
     std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
     if (!sd_cpp_extra_args_str.empty() && !request.gen_params.from_json_str(sd_cpp_extra_args_str)) {
         error_message = "invalid sd_cpp_extra_args";
@@ -228,7 +247,7 @@ static bool build_sdapi_img_gen_request(const json& j,
     }
 
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid params";
         return false;
     }
@@ -347,6 +366,52 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         res.set_content(result.dump(), "application/json");
     });
 
+    svr.Get("/sdapi/v1/upscalers", [runtime](const httplib::Request&, httplib::Response& res) {
+        refresh_upscaler_cache(*runtime);
+
+        auto make_builtin = [](const char* name) {
+            json item;
+            item["name"]       = name;
+            item["model_name"] = nullptr;
+            item["model_path"] = nullptr;
+            item["model_url"]  = nullptr;
+            item["scale"]      = 4;
+            return item;
+        };
+
+        json result = json::array();
+        result.push_back(make_builtin("None"));
+        result.push_back(make_builtin("Lanczos"));
+        result.push_back(make_builtin("Nearest"));
+
+        {
+            std::lock_guard<std::mutex> lock(*runtime->upscaler_mutex);
+            for (const auto& e : *runtime->upscaler_cache) {
+                json item;
+                item["name"]       = e.name;
+                item["model_name"] = e.model_name;
+                item["model_path"] = e.fullpath;
+                item["model_url"]  = nullptr;
+                item["scale"]      = e.scale;
+                result.push_back(item);
+            }
+        }
+
+        res.set_content(result.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/latent-upscale-modes", [](const httplib::Request&, httplib::Response& res) {
+        json result = json::array({
+            {{"name", "Latent"}},
+            {{"name", "Latent (nearest)"}},
+            {{"name", "Latent (nearest-exact)"}},
+            {{"name", "Latent (antialiased)"}},
+            {{"name", "Latent (bicubic)"}},
+            {{"name", "Latent (bicubic antialiased)"}},
+        });
+        res.set_content(result.dump(), "application/json");
+    });
+
     svr.Get("/sdapi/v1/samplers", [runtime](const httplib::Request&, httplib::Response& res) {
         std::vector<std::string> sampler_names;
         sampler_names.push_back("default");

examples/server/routes_sdcpp.cpp

@@ -114,6 +114,17 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
         {"increase_ref_index", defaults.increase_ref_index},
         {"control_strength", defaults.control_strength},
         {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
+        {"hires",
+         {
+             {"enabled", defaults.hires_enabled},
+             {"upscaler", defaults.hires_upscaler},
+             {"scale", defaults.hires_scale},
+             {"target_width", defaults.hires_width},
+             {"target_height", defaults.hires_height},
+             {"steps", defaults.hires_steps},
+             {"denoising_strength", defaults.hires_denoising_strength},
+             {"upscale_tile_size", defaults.hires_upscale_tile_size},
+         }},
         {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
         {"cache_mode", defaults.cache_mode},
         {"cache_option", defaults.cache_option},
@@ -157,6 +168,7 @@ static json make_img_gen_features_json() {
         {"ref_images", true},
         {"lora", true},
         {"vae_tiling", true},
+        {"hires", true},
         {"cache", true},
         {"cancel_queued", true},
         {"cancel_generating", false},
@@ -179,6 +191,7 @@ static json make_vid_gen_features_json() {
 
 static json make_capabilities_json(ServerRuntime& runtime) {
     refresh_lora_cache(runtime);
+    refresh_upscaler_cache(runtime);
 
     AsyncJobManager& manager  = *runtime.async_job_manager;
     const auto& defaults      = *runtime.default_gen_params;
@@ -190,6 +203,7 @@ static json make_capabilities_json(ServerRuntime& runtime) {
     json image_output_formats = supported_img_output_formats();
     json video_output_formats = supported_vid_output_formats();
     json available_loras      = json::array();
+    json available_upscalers  = json::array();
     json supported_modes      = json::array();
 
     for (int i = 0; i < SAMPLE_METHOD_COUNT; ++i) {
@@ -210,6 +224,42 @@ static json make_capabilities_json(ServerRuntime& runtime) {
         }
     }
 
+    available_upscalers.push_back({
+        {"name", "None"},
+    });
+    available_upscalers.push_back({
+        {"name", "Lanczos"},
+    });
+    available_upscalers.push_back({
+        {"name", "Nearest"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (nearest)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (nearest-exact)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (antialiased)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (bicubic)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (bicubic antialiased)"},
+    });
+    {
+        std::lock_guard<std::mutex> lock(*runtime.upscaler_mutex);
+        for (const auto& entry : *runtime.upscaler_cache) {
+            available_upscalers.push_back({
+                {"name", entry.name},
+            });
+        }
+    }
+
     if (supports_img) {
         supported_modes.push_back("img_gen");
     }
@@ -284,6 +334,7 @@ static json make_capabilities_json(ServerRuntime& runtime) {
     result["features"]               = top_level_features;
     result["features_by_mode"]       = features_by_mode;
     result["loras"]                  = available_loras;
+    result["upscalers"]              = available_upscalers;
     return result;
 }
 
@@ -307,7 +358,7 @@ static bool parse_img_gen_request(const json& body,
         return false;
     }
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid generation parameters";
         return false;
     }
@@ -334,7 +385,7 @@ static bool parse_vid_gen_request(const json& body,
         return false;
     }
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(VID_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(VID_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid generation parameters";
         return false;
     }

examples/server/runtime.cpp

@@ -1,6 +1,7 @@
 #include "runtime.h"
 
 #include <algorithm>
+#include <cctype>
 #include <chrono>
 #include <cstdlib>
 #include <filesystem>
@@ -13,6 +14,18 @@
 
 namespace fs = std::filesystem;
 
+static std::string lower_ascii(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
+static bool is_supported_model_ext(const fs::path& p) {
+    auto ext = lower_ascii(p.extension().string());
+    return ext == ".gguf" || ext == ".pt" || ext == ".pth" || ext == ".safetensors";
+}
+
 static const std::string k_base64_chars =
     "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
     "abcdefghijklmnopqrstuvwxyz"
@@ -241,20 +254,12 @@ void refresh_lora_cache(ServerRuntime& rt) {
 
     fs::path lora_dir = rt.ctx_params->lora_model_dir;
     if (fs::exists(lora_dir) && fs::is_directory(lora_dir)) {
-        auto is_lora_ext = [](const fs::path& p) {
-            auto ext = p.extension().string();
-            std::transform(ext.begin(), ext.end(), ext.begin(), [](unsigned char c) {
-                return static_cast<char>(std::tolower(c));
-            });
-            return ext == ".gguf" || ext == ".pt" || ext == ".pth" || ext == ".safetensors";
-        };
-
         for (auto& entry : fs::recursive_directory_iterator(lora_dir)) {
             if (!entry.is_regular_file()) {
                 continue;
             }
             const fs::path& p = entry.path();
-            if (!is_lora_ext(p)) {
+            if (!is_supported_model_ext(p)) {
                 continue;
             }
 
@@ -286,6 +291,40 @@ std::string get_lora_full_path(ServerRuntime& rt, const std::string& path) {
     return it != rt.lora_cache->end() ? it->fullpath : "";
 }
 
+void refresh_upscaler_cache(ServerRuntime& rt) {
+    std::vector<UpscalerEntry> new_cache;
+
+    fs::path upscaler_dir = rt.ctx_params->hires_upscalers_dir;
+    if (fs::exists(upscaler_dir) && fs::is_directory(upscaler_dir)) {
+        for (auto& entry : fs::directory_iterator(upscaler_dir)) {
+            if (!entry.is_regular_file()) {
+                continue;
+            }
+            const fs::path& p = entry.path();
+            if (!is_supported_model_ext(p)) {
+                continue;
+            }
+
+            UpscalerEntry upscaler_entry;
+            upscaler_entry.name       = p.stem().u8string();
+            upscaler_entry.fullpath   = fs::absolute(p).lexically_normal().u8string();
+            upscaler_entry.model_name = "ESRGAN_4x";
+            upscaler_entry.path       = p.filename().u8string();
+
+            new_cache.push_back(std::move(upscaler_entry));
+        }
+    }
+
+    std::sort(new_cache.begin(), new_cache.end(), [](const UpscalerEntry& a, const UpscalerEntry& b) {
+        return a.name < b.name;
+    });
+
+    {
+        std::lock_guard<std::mutex> lock(*rt.upscaler_mutex);
+        *rt.upscaler_cache = std::move(new_cache);
+    }
+}
+
 int64_t unix_timestamp_now() {
     return std::chrono::duration_cast<std::chrono::seconds>(
                std::chrono::system_clock::now().time_since_epoch())

examples/server/runtime.h

@@ -37,6 +37,14 @@ struct LoraEntry {
     std::string fullpath;
 };
 
+struct UpscalerEntry {
+    std::string name;
+    std::string path;
+    std::string fullpath;
+    std::string model_name;
+    int scale = 4;
+};
+
 struct ServerRuntime {
     sd_ctx_t* sd_ctx;
     std::mutex* sd_ctx_mutex;
@@ -45,6 +53,8 @@ struct ServerRuntime {
     const SDGenerationParams* default_gen_params;
     std::vector<LoraEntry>* lora_cache;
     std::mutex* lora_mutex;
+    std::vector<UpscalerEntry>* upscaler_cache;
+    std::mutex* upscaler_mutex;
     AsyncJobManager* async_job_manager;
 };
 
@@ -86,4 +96,5 @@ bool runtime_supports_generation_mode(const ServerRuntime& runtime, SDMode mode)
 std::string unsupported_generation_mode_error(SDMode mode);
 void refresh_lora_cache(ServerRuntime& rt);
 std::string get_lora_full_path(ServerRuntime& rt, const std::string& path);
+void refresh_upscaler_cache(ServerRuntime& rt);
 int64_t unix_timestamp_now();

include/stable-diffusion.h

@@ -203,6 +203,7 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
+    float max_vram;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -289,6 +290,32 @@ typedef struct {
     const char* path;
 } sd_lora_t;
 
+enum sd_hires_upscaler_t {
+    SD_HIRES_UPSCALER_NONE,
+    SD_HIRES_UPSCALER_LATENT,
+    SD_HIRES_UPSCALER_LATENT_NEAREST,
+    SD_HIRES_UPSCALER_LATENT_NEAREST_EXACT,
+    SD_HIRES_UPSCALER_LATENT_ANTIALIASED,
+    SD_HIRES_UPSCALER_LATENT_BICUBIC,
+    SD_HIRES_UPSCALER_LATENT_BICUBIC_ANTIALIASED,
+    SD_HIRES_UPSCALER_LANCZOS,
+    SD_HIRES_UPSCALER_NEAREST,
+    SD_HIRES_UPSCALER_MODEL,
+    SD_HIRES_UPSCALER_COUNT,
+};
+
+typedef struct {
+    bool enabled;
+    enum sd_hires_upscaler_t upscaler;
+    const char* model_path;
+    float scale;
+    int target_width;
+    int target_height;
+    int steps;
+    float denoising_strength;
+    int upscale_tile_size;
+} sd_hires_params_t;
+
 typedef struct {
     const sd_lora_t* loras;
     uint32_t lora_count;
@@ -312,6 +339,7 @@ typedef struct {
     sd_pm_params_t pm_params;
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
+    sd_hires_params_t hires;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -365,8 +393,11 @@ SD_API const char* sd_preview_name(enum preview_t preview);
 SD_API enum preview_t str_to_preview(const char* str);
 SD_API const char* sd_lora_apply_mode_name(enum lora_apply_mode_t mode);
 SD_API enum lora_apply_mode_t str_to_lora_apply_mode(const char* str);
+SD_API const char* sd_hires_upscaler_name(enum sd_hires_upscaler_t upscaler);
+SD_API enum sd_hires_upscaler_t str_to_sd_hires_upscaler(const char* str);
 
 SD_API void sd_cache_params_init(sd_cache_params_t* cache_params);
+SD_API void sd_hires_params_init(sd_hires_params_t* hires_params);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);

src/anima.hpp

@@ -499,9 +499,15 @@ namespace Anima {
                 encoder_hidden_states = adapted_context;
             }
 
+            sd::ggml_graph_cut::mark_graph_cut(x, "anima.prelude", "x");
+            sd::ggml_graph_cut::mark_graph_cut(embedded_timestep, "anima.prelude", "embedded_timestep");
+            sd::ggml_graph_cut::mark_graph_cut(temb, "anima.prelude", "temb");
+            sd::ggml_graph_cut::mark_graph_cut(encoder_hidden_states, "anima.prelude", "context");
+
             for (int i = 0; i < num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["blocks." + std::to_string(i)]);
                 x          = block->forward(ctx, x, encoder_hidden_states, embedded_timestep, temb, image_pe);
+                sd::ggml_graph_cut::mark_graph_cut(x, "anima.blocks." + std::to_string(i), "x");
             }
 
             x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]

src/auto_encoder_kl.hpp

@@ -328,6 +328,7 @@ public:
         auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
 
         auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.prelude", "h");
 
         // downsampling
         size_t num_resolutions = ch_mult.size();
@@ -337,12 +338,14 @@ public:
                 auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
 
                 h = down_block->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.down." + std::to_string(i) + ".block." + std::to_string(j), "h");
             }
             if (i != num_resolutions - 1) {
                 std::string name = "down." + std::to_string(i) + ".downsample";
                 auto down_sample = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
 
                 h = down_sample->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.down." + std::to_string(i) + ".downsample", "h");
             }
         }
 
@@ -350,6 +353,7 @@ public:
         h = mid_block_1->forward(ctx, h);
         h = mid_attn_1->forward(ctx, h);
         h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.mid", "h");
 
         // end
         h = norm_out->forward(ctx, h);
@@ -450,6 +454,7 @@ public:
 
         // conv_in
         auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.prelude", "h");
 
         // middle
         h = mid_block_1->forward(ctx, h);
@@ -457,6 +462,7 @@ public:
 
         h = mid_attn_1->forward(ctx, h);
         h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.mid", "h");
 
         // upsampling
         int num_resolutions = static_cast<int>(ch_mult.size());
@@ -466,12 +472,14 @@ public:
                 auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
 
                 h = up_block->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.up." + std::to_string(i) + ".block." + std::to_string(j), "h");
             }
             if (i != 0) {
                 std::string name = "up." + std::to_string(i) + ".upsample";
                 auto up_sample   = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
 
                 h = up_sample->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.up." + std::to_string(i) + ".upsample", "h");
             }
         }
 
@@ -599,6 +607,7 @@ public:
         if (use_quant) {
             auto post_quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["post_quant_conv"]);
             z                    = post_quant_conv->forward(ctx, z);  // [N, z_channels, h, w]
+            // sd::ggml_graph_cut::mark_graph_cut(z, "vae.decode.prelude", "z");
         }
         auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
 
@@ -616,6 +625,7 @@ public:
         if (use_quant) {
             auto quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["quant_conv"]);
             z               = quant_conv->forward(ctx, z);  // [N, 2*embed_dim, h/8, w/8]
+            // sd::ggml_graph_cut::mark_graph_cut(z, "vae.encode.final", "z");
         }
         if (sd_version_uses_flux2_vae(version)) {
             z = ggml_ext_chunk(ctx->ggml_ctx, z, 2, 2)[0];

src/clip.hpp

@@ -96,7 +96,8 @@ public:
     ggml_tensor* forward(GGMLRunnerContext* ctx,
                          ggml_tensor* x,
                          ggml_tensor* mask                   = nullptr,
-                         int clip_skip     = -1) {
+                         int clip_skip                       = -1,
+                         const std::string& graph_cut_prefix = "") {
         // x: [N, n_token, d_model]
         int layer_idx = n_layer - 1;
         // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -112,6 +113,9 @@ public:
             std::string name = "layers." + std::to_string(i);
             auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
             x                = layer->forward(ctx, x, mask);  // [N, n_token, d_model]
+            if (!graph_cut_prefix.empty()) {
+                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".layers." + std::to_string(i), "x");
+            }
             // LOG_DEBUG("layer %d", i);
         }
         return x;
@@ -304,7 +308,8 @@ public:
         auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
 
         auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
-        x      = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip);
+        sd::ggml_graph_cut::mark_graph_cut(x, "clip_text.prelude", "x");
+        x = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip, "clip_text");
         if (return_pooled || with_final_ln) {
             x = final_layer_norm->forward(ctx, x);
         }
@@ -368,7 +373,8 @@ public:
 
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
-        x      = encoder->forward(ctx, x, nullptr, clip_skip);
+        sd::ggml_graph_cut::mark_graph_cut(x, "clip_vision.prelude", "x");
+        x = encoder->forward(ctx, x, nullptr, clip_skip, "clip_vision");
 
         auto last_hidden_state = x;
 

src/common_block.hpp

@@ -1,7 +1,9 @@
 #ifndef __COMMON_BLOCK_HPP__
 #define __COMMON_BLOCK_HPP__
 
+#include "ggml-backend.h"
 #include "ggml_extend.hpp"
+#include "util.h"
 
 class DownSampleBlock : public GGMLBlock {
 protected:
@@ -248,9 +250,6 @@ public:
         float scale         = 1.f;
         if (precision_fix) {
             scale = 1.f / 128.f;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
         }
         // The purpose of the scale here is to prevent NaN issues in certain situations.
         // For example, when using Vulkan without enabling force_prec_f32,
@@ -264,6 +263,9 @@ public:
 
         auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
         auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
+        if (sd_backend_is(ctx->backend, "Vulkan")) {
+            net_2->set_force_prec_f32(true);
+        }
 
         x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
         x = net_2->forward(ctx, x);  // [ne3, ne2, ne1, dim_out]

src/conditioner.hpp

@@ -85,6 +85,7 @@ public:
     virtual void free_params_buffer()                                                      = 0;
     virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)           = 0;
     virtual size_t get_params_buffer_size()                                                = 0;
+    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
     virtual void set_flash_attention_enabled(bool enabled) = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(int n_threads,
@@ -165,6 +166,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        text_model->set_max_graph_vram_bytes(max_vram_bytes);
+        if (sd_version_is_sdxl(version)) {
+            text_model2->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         text_model->set_flash_attention_enabled(enabled);
         if (sd_version_is_sdxl(version)) {
@@ -781,6 +789,18 @@ struct SD3CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (clip_l) {
+            clip_l->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+        if (clip_g) {
+            clip_g->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1124,6 +1144,15 @@ struct FluxCLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (clip_l) {
+            clip_l->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1349,6 +1378,12 @@ struct T5CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (t5) {
             t5->set_flash_attention_enabled(enabled);
@@ -1525,6 +1560,10 @@ struct AnimaConditioner : public Conditioner {
         return llm->get_params_buffer_size();
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        llm->set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -1657,6 +1696,10 @@ struct LLMEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        llm->set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }

src/denoiser.hpp

@@ -808,6 +808,18 @@ static std::tuple<float, float, float> get_ancestral_step_flow(float sigma_from,
     return {sigma_down, sigma_up, alpha_scale};
 }
 
+static std::tuple<float, float, float> get_ancestral_step(float sigma_from,
+                                                          float sigma_to,
+                                                          float eta,
+                                                          bool is_flow_denoiser) {
+    if (is_flow_denoiser) {
+        return get_ancestral_step_flow(sigma_from, sigma_to, eta);
+    } else {
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
+        return {sigma_down, sigma_up, 1.0f};
+    }
+}
+
 static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
                                                 sd::Tensor<float> x,
                                                 const std::vector<float>& sigmas,
@@ -1247,6 +1259,7 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
                                               sd::Tensor<float> x,
                                               const std::vector<float>& sigmas,
                                               std::shared_ptr<RNG> rng,
+                                              bool is_flow_denoiser,
                                               float eta) {
     sd::Tensor<float> old_denoised = x;
     bool have_old_sigma            = false;
@@ -1278,7 +1291,8 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
 
         float sigma_from            = sigmas[i];
         float sigma_to              = sigmas[i + 1];
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
+
+        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);
 
         if (sigma_down == 0.0f || !have_old_sigma) {
             x += ((x - denoised) / sigma_from) * (sigma_down - sigma_from);
@@ -1305,7 +1319,10 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
             x = sigma_fn(h) * x + h * (b1 * denoised + b2 * old_denoised);
         }
 
-        if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
+        if (sigma_to > 0.0f && sigma_up > 0.0f) {
+            if (is_flow_denoiser) {
+                x *= alpha_scale;
+            }
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
 
@@ -1320,6 +1337,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
                                        sd::Tensor<float> x,
                                        const std::vector<float>& sigmas,
                                        std::shared_ptr<RNG> rng,
+                                       bool is_flow_denoiser,
                                        float eta) {
     const float c2 = 0.5f;
     auto t_fn      = [](float sigma) -> float { return -logf(sigma); };
@@ -1348,7 +1366,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
         }
         sd::Tensor<float> denoised = std::move(denoised_opt);
 
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
+        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);
 
         sd::Tensor<float> x0 = x;
         if (sigma_down == 0.0f || sigma_from == 0.0f) {
@@ -1377,7 +1395,10 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
             x                           = x0 + h * (b1 * eps1 + b2 * eps2);
         }
 
-        if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
+        if (sigma_to > 0.0f && sigma_up > 0.0f) {
+            if (is_flow_denoiser) {
+                x *= alpha_scale;
+            }
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
     }
@@ -1664,9 +1685,9 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case IPNDM_V_SAMPLE_METHOD:
             return sample_ipndm_v(model, std::move(x), sigmas);
         case RES_MULTISTEP_SAMPLE_METHOD:
-            return sample_res_multistep(model, std::move(x), sigmas, rng, eta);
+            return sample_res_multistep(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case RES_2S_SAMPLE_METHOD:
-            return sample_res_2s(model, std::move(x), sigmas, rng, eta);
+            return sample_res_2s(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case ER_SDE_SAMPLE_METHOD:
             return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case DDIM_TRAILING_SAMPLE_METHOD:

src/diffusion_model.hpp

@@ -49,6 +49,7 @@ struct DiffusionModel {
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
     virtual int64_t get_adm_in_channels()                            = 0;
     virtual void set_flash_attention_enabled(bool enabled)           = 0;
+    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes)     = 0;
     virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };
 
@@ -98,6 +99,10 @@ struct UNetModel : public DiffusionModel {
         unet.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        unet.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         unet.set_circular_axes(circular_x, circular_y);
     }
@@ -164,6 +169,10 @@ struct MMDiTModel : public DiffusionModel {
         mmdit.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        mmdit.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         mmdit.set_circular_axes(circular_x, circular_y);
     }
@@ -229,6 +238,10 @@ struct FluxModel : public DiffusionModel {
         flux.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        flux.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         flux.set_circular_axes(circular_x, circular_y);
     }
@@ -299,6 +312,10 @@ struct AnimaModel : public DiffusionModel {
         anima.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        anima.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         anima.set_circular_axes(circular_x, circular_y);
     }
@@ -364,6 +381,10 @@ struct WanModel : public DiffusionModel {
         wan.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        wan.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         wan.set_circular_axes(circular_x, circular_y);
     }
@@ -433,6 +454,10 @@ struct QwenImageModel : public DiffusionModel {
         qwen_image.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        qwen_image.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         qwen_image.set_circular_axes(circular_x, circular_y);
     }
@@ -499,6 +524,10 @@ struct ZImageModel : public DiffusionModel {
         z_image.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        z_image.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         z_image.set_circular_axes(circular_x, circular_y);
     }
@@ -564,6 +593,10 @@ struct ErnieImageModel : public DiffusionModel {
         ernie_image.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        ernie_image.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         ernie_image.set_circular_axes(circular_x, circular_y);
     }

src/ernie_image.hpp

@@ -295,6 +295,8 @@ namespace ErnieImage {
             auto c      = time_embedding->forward(ctx, sample);  // [N, hidden_size]
 
             auto mod_params = adaLN_mod->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, 6 * hidden_size]
+            sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.prelude", "hidden_states");
+            // sd::ggml_graph_cut::mark_graph_cut(mod_params, "ernie_image.prelude", "mod_params");
             auto chunks = ggml_ext_chunk(ctx->ggml_ctx, mod_params, 6, 0);
             std::vector<ggml_tensor*> temb;
             temb.reserve(6);
@@ -305,6 +307,7 @@ namespace ErnieImage {
             for (int i = 0; i < params.num_layers; i++) {
                 auto layer    = std::dynamic_pointer_cast<ErnieImageSharedAdaLNBlock>(blocks["layers." + std::to_string(i)]);
                 hidden_states = layer->forward(ctx, hidden_states, pe, temb);
+                sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.layers." + std::to_string(i), "hidden_states");
             }
 
             hidden_states = final_norm->forward(ctx, hidden_states, c);

src/esrgan.hpp

@@ -125,26 +125,32 @@ public:
         auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
 
         auto feat = conv_first->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.prelude", "feat");
         auto body_feat = feat;
         for (int i = 0; i < num_block; i++) {
             std::string name = "body." + std::to_string(i);
             auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
 
             body_feat = block->forward(ctx, body_feat);
+            sd::ggml_graph_cut::mark_graph_cut(body_feat, "esrgan.body." + std::to_string(i), "feat");
         }
         body_feat = conv_body->forward(ctx, body_feat);
         feat      = ggml_add(ctx->ggml_ctx, feat, body_feat);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.body.out", "feat");
         // upsample
         if (scale >= 2) {
             auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
             feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up1", "feat");
             if (scale == 4) {
                 auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
                 feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+                sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up2", "feat");
             }
         }
         // for all scales
         auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        sd::ggml_graph_cut::mark_graph_cut(out, "esrgan.final", "out");
         return out;
     }
 };

src/flux.hpp

@@ -928,6 +928,9 @@ namespace Flux {
             }
 
             txt = txt_in->forward(ctx, txt);
+            sd::ggml_graph_cut::mark_graph_cut(img, "flux.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(txt, "flux.prelude", "txt");
+            sd::ggml_graph_cut::mark_graph_cut(vec, "flux.prelude", "vec");
 
             for (int i = 0; i < params.depth; i++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
@@ -939,6 +942,8 @@ namespace Flux {
                 auto img_txt = block->forward(ctx, img, txt, vec, pe, txt_img_mask, ds_img_mods, ds_txt_mods);
                 img          = img_txt.first;   // [N, n_img_token, hidden_size]
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
+                sd::ggml_graph_cut::mark_graph_cut(img, "flux.double_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "flux.double_blocks." + std::to_string(i), "txt");
             }
 
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
@@ -949,6 +954,7 @@ namespace Flux {
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask, ss_mods);
+                sd::ggml_graph_cut::mark_graph_cut(txt_img, "flux.single_blocks." + std::to_string(i), "txt_img");
             }
 
             img = ggml_view_3d(ctx->ggml_ctx,

src/ggml_extend_backend.hpp

@@ -0,0 +1,298 @@
+#ifndef __GGML_EXTEND_BACKEND_HPP__
+#define __GGML_EXTEND_BACKEND_HPP__
+
+#include <cstring>
+#include <mutex>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifndef __STATIC_INLINE__
+#define __STATIC_INLINE__ static inline
+#endif
+
+inline void ggml_backend_load_all_once() {
+    // If the registry already has devices and the CPU backend is present,
+    // assume either static registration or explicit host-side preloading has
+    // completed and avoid rescanning the default paths.
+    if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+        return;
+    }
+    // In dynamic-backend mode the backend modules are discovered at runtime,
+    // so we must load them before asking for the CPU backend or its proc table.
+    // If the host preloaded only a subset of backends, allow one default-path
+    // scan so missing modules can still be discovered.
+    static std::once_flag once;
+    std::call_once(once, []() {
+        if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+            return;
+        }
+        ggml_backend_load_all();
+    });
+}
+
+// Do not gate this branch on GGML_CPU or GGML_CPU_ALL_VARIANTS:
+// those are CMake options used to configure ggml itself, but they are not
+// exported as PUBLIC compile definitions to stable-diffusion in backend-DL mode.
+// In practice, this target can reliably see GGML_BACKEND_DL, but not whether
+// the CPU backend was compiled as a loadable module. We therefore use runtime
+// backend discovery instead of compile-time assumptions.
+
+__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_cpu_reg() {
+    ggml_backend_reg_t reg = ggml_backend_reg_by_name("CPU");
+    if (reg != nullptr) {
+        return reg;
+    }
+
+    ggml_backend_load_all_once();
+    return ggml_backend_reg_by_name("CPU");
+}
+
+__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_reg_from_backend(ggml_backend_t backend) {
+    if (backend != nullptr) {
+        ggml_backend_dev_t device = ggml_backend_get_device(backend);
+        if (device != nullptr) {
+            return ggml_backend_dev_backend_reg(device);
+        }
+    }
+
+    return ggml_backend_cpu_reg();
+}
+
+__STATIC_INLINE__ ggml_backend_t ggml_backend_cpu_init() {
+    ggml_backend_t backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+    if (backend != nullptr) {
+        return backend;
+    }
+
+    ggml_backend_load_all_once();
+    return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+}
+
+__STATIC_INLINE__ bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    if (backend == nullptr) {
+        return false;
+    }
+
+    ggml_backend_dev_t device = ggml_backend_get_device(backend);
+    if (device != nullptr) {
+        return ggml_backend_dev_type(device) == GGML_BACKEND_DEVICE_TYPE_CPU;
+    }
+
+    const char* backend_name = ggml_backend_name(backend);
+    return backend_name != nullptr && std::strcmp(backend_name, "CPU") == 0;
+}
+
+__STATIC_INLINE__ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
+    if (reg == nullptr) {
+        return;
+    }
+
+    auto fn = reinterpret_cast<ggml_backend_set_n_threads_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"));
+    if (fn != nullptr) {
+        fn(backend_cpu, n_threads);
+    }
+}
+
+using __ggml_backend_cpu_set_threadpool_t = void (*)(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+
+__STATIC_INLINE__ void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
+    if (reg == nullptr) {
+        return;
+    }
+
+    auto fn = reinterpret_cast<__ggml_backend_cpu_set_threadpool_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"));
+    if (fn != nullptr) {
+        fn(backend_cpu, threadpool);
+    }
+}
+
+__STATIC_INLINE__ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void* abort_callback_data) {
+    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
+    if (reg == nullptr) {
+        return;
+    }
+
+    auto fn = reinterpret_cast<ggml_backend_set_abort_callback_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"));
+    if (fn != nullptr) {
+        fn(backend_cpu, abort_callback, abort_callback_data);
+    }
+}
+
+__STATIC_INLINE__ ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr) {
+        return nullptr;
+    }
+
+    return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+}
+
+__STATIC_INLINE__ bool ggml_backend_tensor_is_host_accessible(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr || tensor->data == nullptr) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_tensor_buffer(tensor);
+    return buffer == nullptr || ggml_backend_buffer_is_host(buffer);
+}
+
+__STATIC_INLINE__ size_t ggml_backend_tensor_offset(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+    return (size_t)(i0 * tensor->nb[0] + i1 * tensor->nb[1] + i2 * tensor->nb[2] + i3 * tensor->nb[3]);
+}
+
+template <typename T>
+__STATIC_INLINE__ void ggml_backend_tensor_write_scalar(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, T value) {
+    const size_t offset = ggml_backend_tensor_offset(tensor, i0, i1, i2, i3);
+
+    if (ggml_backend_tensor_is_host_accessible(tensor)) {
+        auto* dst = reinterpret_cast<T*>(reinterpret_cast<char*>(tensor->data) + offset);
+        *dst      = value;
+        return;
+    }
+
+    ggml_backend_tensor_set(const_cast<struct ggml_tensor*>(tensor), &value, offset, sizeof(T));
+}
+
+__STATIC_INLINE__ void ggml_set_f32_nd(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float value) {
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+__STATIC_INLINE__ void ggml_set_f32_1d(const struct ggml_tensor* tensor, int i, float value) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = {0, 0, 0, 0};
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
+        return;
+    }
+
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+__STATIC_INLINE__ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context* ctx, struct ggml_cgraph* cgraph, int n_threads) {
+    (void)ctx;
+
+    // The legacy ggml_graph_compute_with_ctx() symbol lives in ggml-cpu, but
+    // the backend proc table does not expose it in GGML_BACKEND_DL mode.
+    // Recreate the old behavior by initializing the CPU backend explicitly and
+    // executing the graph through the generic backend API.
+    ggml_backend_t backend = ggml_backend_cpu_init();
+    if (backend == nullptr) {
+        return GGML_STATUS_ALLOC_FAILED;
+    }
+
+    ggml_backend_cpu_set_n_threads(backend, n_threads);
+
+    const enum ggml_status status = ggml_backend_graph_compute(backend, cgraph);
+    ggml_backend_free(backend);
+
+    return status;
+}
+
+__STATIC_INLINE__ ggml_tensor* ggml_set_f32(struct ggml_tensor* tensor, float value) {
+    GGML_ASSERT(tensor != nullptr);
+
+    if (ggml_backend_tensor_is_host_accessible(tensor) && ggml_is_contiguous(tensor)) {
+        const int64_t nelements = ggml_nelements(tensor);
+
+        switch (tensor->type) {
+            case GGML_TYPE_I8: {
+                auto* data     = reinterpret_cast<int8_t*>(tensor->data);
+                const int8_t v = static_cast<int8_t>(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_I16: {
+                auto* data      = reinterpret_cast<int16_t*>(tensor->data);
+                const int16_t v = static_cast<int16_t>(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_I32: {
+                auto* data      = reinterpret_cast<int32_t*>(tensor->data);
+                const int32_t v = static_cast<int32_t>(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_F16: {
+                auto* data          = reinterpret_cast<ggml_fp16_t*>(tensor->data);
+                const ggml_fp16_t v = ggml_fp32_to_fp16(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_BF16: {
+                auto* data          = reinterpret_cast<ggml_bf16_t*>(tensor->data);
+                const ggml_bf16_t v = ggml_fp32_to_bf16(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_F32: {
+                auto* data = reinterpret_cast<float*>(tensor->data);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = value;
+                }
+            } break;
+            default:
+                GGML_ABORT("fatal error");
+        }
+
+        return tensor;
+    }
+
+    const int64_t nelements = ggml_nelements(tensor);
+    for (int64_t i = 0; i < nelements; ++i) {
+        ggml_set_f32_1d(tensor, static_cast<int>(i), value);
+    }
+
+    return tensor;
+}
+
+#endif

src/ggml_graph_cut.cpp

@@ -0,0 +1,676 @@
+#include "ggml_graph_cut.h"
+
+#include <algorithm>
+#include <cstring>
+#include <map>
+#include <set>
+#include <sstream>
+#include <stack>
+#include <unordered_map>
+
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "util.h"
+
+#include "../ggml/src/ggml-impl.h"
+
+namespace sd::ggml_graph_cut {
+
+    static std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return "<null>";
+        }
+        if (tensor->name[0] != '\0') {
+            return tensor->name;
+        }
+        return sd_format("<tensor@%p>", (const void*)tensor);
+    }
+
+    static int graph_leaf_index(ggml_cgraph* gf, const ggml_tensor* tensor) {
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(tensor != nullptr);
+        for (int i = 0; i < gf->n_leafs; ++i) {
+            if (gf->leafs[i] == tensor) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    static bool is_params_tensor(const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                                 const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return false;
+        }
+        return params_tensor_set.find(tensor) != params_tensor_set.end();
+    }
+
+    static Plan::InputShape input_shape(const ggml_tensor* tensor) {
+        Plan::InputShape shape;
+        if (tensor == nullptr) {
+            return shape;
+        }
+        shape.type = tensor->type;
+        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+            shape.ne[static_cast<size_t>(i)] = tensor->ne[i];
+        }
+        return shape;
+    }
+
+    static size_t graph_cut_segment_vram_bytes(const Segment& segment) {
+        return segment.compute_buffer_size +
+               segment.input_param_bytes +
+               segment.input_previous_cut_bytes +
+               segment.output_bytes;
+    }
+
+    static Segment make_segment_seed(const Plan& plan,
+                                     size_t start_segment_index,
+                                     size_t end_segment_index) {
+        GGML_ASSERT(start_segment_index < plan.segments.size());
+        GGML_ASSERT(end_segment_index < plan.segments.size());
+        GGML_ASSERT(start_segment_index <= end_segment_index);
+
+        Segment seed;
+        const auto& start_segment  = plan.segments[start_segment_index];
+        const auto& target_segment = plan.segments[end_segment_index];
+        std::unordered_set<int> seen_output_node_indices;
+        for (size_t seg_idx = start_segment_index; seg_idx <= end_segment_index; ++seg_idx) {
+            for (int output_node_index : plan.segments[seg_idx].output_node_indices) {
+                if (seen_output_node_indices.insert(output_node_index).second) {
+                    seed.output_node_indices.push_back(output_node_index);
+                }
+            }
+        }
+        if (start_segment_index == end_segment_index) {
+            seed.group_name = target_segment.group_name;
+        } else {
+            seed.group_name = sd_format("%s..%s",
+                                        start_segment.group_name.c_str(),
+                                        target_segment.group_name.c_str());
+        }
+        return seed;
+    }
+
+    static void build_segment(ggml_cgraph* gf,
+                              Plan& plan,
+                              Segment& segment,
+                              const std::unordered_map<const ggml_tensor*, int>& producer_index,
+                              std::unordered_set<int>& available_cut_output_node_indices,
+                              ggml_backend_t backend,
+                              const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                              const char* log_desc) {
+        std::set<int> internal_nodes;
+        std::unordered_set<const ggml_tensor*> input_seen;
+        std::vector<Segment::InputRef> input_refs;
+
+        std::stack<ggml_tensor*> work_stack;
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output != nullptr) {
+                work_stack.push(output);
+            }
+        }
+
+        while (!work_stack.empty()) {
+            ggml_tensor* tensor = work_stack.top();
+            work_stack.pop();
+
+            if (tensor == nullptr) {
+                continue;
+            }
+
+            auto producer_it = producer_index.find(tensor);
+            if (producer_it == producer_index.end()) {
+                if (input_seen.insert(tensor).second) {
+                    Segment::InputRef input_ref;
+                    input_ref.type         = is_params_tensor(params_tensor_set, tensor) ? Segment::INPUT_PARAM : Segment::INPUT_EXTERNAL;
+                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
+                    input_ref.leaf_index   = graph_leaf_index(gf, tensor);
+                    input_refs.push_back(std::move(input_ref));
+                }
+                continue;
+            }
+
+            int node_idx = producer_it->second;
+            if (available_cut_output_node_indices.find(node_idx) != available_cut_output_node_indices.end()) {
+                if (input_seen.insert(tensor).second) {
+                    Segment::InputRef input_ref;
+                    input_ref.type         = Segment::INPUT_PREVIOUS_CUT;
+                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
+                    input_ref.node_index   = node_idx;
+                    input_refs.push_back(std::move(input_ref));
+                }
+                continue;
+            }
+
+            if (!internal_nodes.insert(node_idx).second) {
+                continue;
+            }
+
+            ggml_tensor* node = ggml_graph_node(gf, node_idx);
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+                if (node->src[src_idx] != nullptr) {
+                    work_stack.push(node->src[src_idx]);
+                }
+            }
+        }
+
+        if (!internal_nodes.empty()) {
+            segment.internal_node_indices.assign(internal_nodes.begin(), internal_nodes.end());
+        }
+
+        std::sort(input_refs.begin(),
+                  input_refs.end(),
+                  [](const Segment::InputRef& a, const Segment::InputRef& b) {
+                      if (a.type != b.type) {
+                          return a.type < b.type;
+                      }
+                      return a.display_name < b.display_name;
+                  });
+        segment.input_refs = input_refs;
+        for (const auto& input : input_refs) {
+            ggml_tensor* current_input = input_tensor(gf, input);
+            size_t tensor_bytes        = current_input == nullptr
+                                             ? 0
+                                             : (input.type == Segment::INPUT_PREVIOUS_CUT
+                                                    ? cache_tensor_bytes(current_input)
+                                                    : ggml_nbytes(current_input));
+            switch (input.type) {
+                case Segment::INPUT_PREVIOUS_CUT:
+                    segment.input_previous_cut_bytes += tensor_bytes;
+                    break;
+                case Segment::INPUT_PARAM:
+                    segment.input_param_bytes += tensor_bytes;
+                    break;
+                case Segment::INPUT_EXTERNAL:
+                default:
+                    segment.input_external_bytes += tensor_bytes;
+                    break;
+            }
+        }
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            segment.output_bytes += cache_tensor_bytes(output);
+        }
+        segment.compute_buffer_size = measure_segment_compute_buffer(backend, gf, segment, log_desc);
+
+        for (int output_node_index : segment.output_node_indices) {
+            available_cut_output_node_indices.insert(output_node_index);
+        }
+        plan.segments.push_back(std::move(segment));
+    }
+
+    bool is_graph_cut_tensor(const ggml_tensor* tensor) {
+        if (tensor == nullptr || tensor->name[0] == '\0') {
+            return false;
+        }
+        return std::strncmp(tensor->name, GGML_RUNNER_CUT_PREFIX, std::strlen(GGML_RUNNER_CUT_PREFIX)) == 0;
+    }
+
+    std::string make_graph_cut_name(const std::string& group, const std::string& output) {
+        return std::string(GGML_RUNNER_CUT_PREFIX) + group + "|" + output;
+    }
+
+    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output) {
+        if (tensor == nullptr) {
+            return;
+        }
+        auto name = make_graph_cut_name(group, output);
+        ggml_set_name(tensor, name.c_str());
+    }
+
+    int leaf_count(ggml_cgraph* gf) {
+        GGML_ASSERT(gf != nullptr);
+        return gf->n_leafs;
+    }
+
+    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index) {
+        GGML_ASSERT(gf != nullptr);
+        if (leaf_index < 0 || leaf_index >= gf->n_leafs) {
+            return nullptr;
+        }
+        return gf->leafs[leaf_index];
+    }
+
+    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return nullptr;
+        }
+        return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    }
+
+    ggml_tensor* cache_source_tensor(ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return nullptr;
+        }
+        return tensor->view_src ? tensor->view_src : tensor;
+    }
+
+    size_t cache_tensor_bytes(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return 0;
+        }
+        const ggml_tensor* cache_src = tensor->view_src ? tensor->view_src : tensor;
+        return ggml_nbytes(cache_src);
+    }
+
+    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) {
+        GGML_ASSERT(gf != nullptr);
+        if (ggml_graph_n_nodes(gf) != plan.n_nodes || gf->n_leafs != plan.n_leafs) {
+            return false;
+        }
+        for (const auto& input_shape_ref : plan.input_shapes) {
+            if (input_shape_ref.leaf_index < 0 || input_shape_ref.leaf_index >= gf->n_leafs) {
+                return false;
+            }
+            ggml_tensor* leaf = gf->leafs[input_shape_ref.leaf_index];
+            if (leaf == nullptr || input_shape_ref.type != leaf->type) {
+                return false;
+            }
+            for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+                if (input_shape_ref.ne[static_cast<size_t>(d)] != leaf->ne[d]) {
+                    return false;
+                }
+            }
+        }
+        return true;
+    }
+
+    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index) {
+        GGML_ASSERT(gf != nullptr);
+        if (output_index >= segment.output_node_indices.size()) {
+            return nullptr;
+        }
+        int node_index = segment.output_node_indices[output_index];
+        if (node_index < 0 || node_index >= ggml_graph_n_nodes(gf)) {
+            return nullptr;
+        }
+        return ggml_graph_node(gf, node_index);
+    }
+
+    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref) {
+        GGML_ASSERT(gf != nullptr);
+        if (input_ref.type == Segment::INPUT_PREVIOUS_CUT) {
+            if (input_ref.node_index < 0 || input_ref.node_index >= ggml_graph_n_nodes(gf)) {
+                return nullptr;
+            }
+            return ggml_graph_node(gf, input_ref.node_index);
+        }
+        if (input_ref.leaf_index < 0 || input_ref.leaf_index >= gf->n_leafs) {
+            return nullptr;
+        }
+        return leaf_tensor(gf, input_ref.leaf_index);
+    }
+
+    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment) {
+        GGML_ASSERT(gf != nullptr);
+        std::vector<ggml_tensor*> tensors;
+        std::unordered_set<ggml_tensor*> seen_tensors;
+        tensors.reserve(segment.input_refs.size());
+        seen_tensors.reserve(segment.input_refs.size());
+        for (const auto& input_ref : segment.input_refs) {
+            if (input_ref.type != Segment::INPUT_PARAM) {
+                continue;
+            }
+            ggml_tensor* tensor = input_tensor(gf, input_ref);
+            if (tensor == nullptr) {
+                continue;
+            }
+            if (seen_tensors.insert(tensor).second) {
+                tensors.push_back(tensor);
+            }
+        }
+        return tensors;
+    }
+
+    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc) {
+        std::vector<ggml_tensor*> tensors = param_tensors(gf, segment);
+        std::vector<ggml_tensor*> filtered_tensors;
+        filtered_tensors.reserve(tensors.size());
+        for (ggml_tensor* tensor : tensors) {
+            if (tensor_buffer(tensor) == nullptr) {
+                LOG_WARN("%s graph cut skipping param input without buffer: segment=%s tensor=%s",
+                         log_desc == nullptr ? "unknown" : log_desc,
+                         segment.group_name.c_str(),
+                         tensor->name);
+                continue;
+            }
+            filtered_tensors.push_back(tensor);
+        }
+        return filtered_tensors;
+    }
+
+    std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
+                                                               const Plan& plan,
+                                                               size_t current_segment_index) {
+        GGML_ASSERT(gf != nullptr);
+        std::unordered_set<std::string> future_input_names;
+        for (size_t seg_idx = current_segment_index + 1; seg_idx < plan.segments.size(); ++seg_idx) {
+            const auto& segment = plan.segments[seg_idx];
+            for (const auto& input_ref : segment.input_refs) {
+                if (input_ref.type != Segment::INPUT_PREVIOUS_CUT) {
+                    continue;
+                }
+                ggml_tensor* current_input = input_tensor(gf, input_ref);
+                if (current_input != nullptr && current_input->name[0] != '\0') {
+                    future_input_names.insert(current_input->name);
+                }
+            }
+        }
+        return future_input_names;
+    }
+
+    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
+                                     const Segment& segment,
+                                     ggml_context** graph_ctx_out) {
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(graph_ctx_out != nullptr);
+
+        const size_t graph_size = segment.internal_node_indices.size() + segment.input_refs.size() + 8;
+        ggml_init_params params = {
+            /*.mem_size   =*/ggml_graph_overhead_custom(graph_size, false) + 1024,
+            /*.mem_buffer =*/nullptr,
+            /*.no_alloc   =*/true,
+        };
+        ggml_context* graph_ctx = ggml_init(params);
+        GGML_ASSERT(graph_ctx != nullptr);
+        ggml_cgraph* segment_graph = ggml_new_graph_custom(graph_ctx, graph_size, false);
+        GGML_ASSERT(segment_graph != nullptr);
+
+        for (const auto& input : segment.input_refs) {
+            ggml_tensor* current_input = input_tensor(gf, input);
+            if (current_input == nullptr) {
+                continue;
+            }
+            GGML_ASSERT(segment_graph->n_leafs < segment_graph->size);
+            segment_graph->leafs[segment_graph->n_leafs++] = current_input;
+        }
+
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output == nullptr) {
+                continue;
+            }
+            ggml_set_output(output);
+        }
+        for (int node_idx : segment.internal_node_indices) {
+            ggml_graph_add_node(segment_graph, ggml_graph_node(gf, node_idx));
+        }
+        *graph_ctx_out = graph_ctx;
+        return segment_graph;
+    }
+
+    size_t measure_segment_compute_buffer(ggml_backend_t backend,
+                                          ggml_cgraph* gf,
+                                          const Segment& segment,
+                                          const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        if (segment.internal_node_indices.empty()) {
+            return 0;
+        }
+
+        ggml_context* graph_ctx    = nullptr;
+        ggml_cgraph* segment_graph = build_segment_graph(gf, segment, &graph_ctx);
+        ggml_gallocr_t allocr      = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
+
+        size_t sizes[1] = {0};
+        ggml_gallocr_reserve_n_size(
+            allocr,
+            segment_graph,
+            nullptr,
+            nullptr,
+            sizes);
+        size_t buffer_size = sizes[0];
+
+        ggml_gallocr_free(allocr);
+        ggml_free(graph_ctx);
+        return buffer_size;
+    }
+
+    Plan build_plan(ggml_backend_t backend,
+                    ggml_cgraph* gf,
+                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                    const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        Plan plan;
+        plan.available    = true;
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        if (n_nodes <= 0) {
+            return plan;
+        }
+        plan.n_nodes = n_nodes;
+        plan.n_leafs = gf->n_leafs;
+        for (int i = 0; i < gf->n_leafs; ++i) {
+            ggml_tensor* leaf = gf->leafs[i];
+            if (is_params_tensor(params_tensor_set, leaf)) {
+                continue;
+            }
+            auto shape       = input_shape(leaf);
+            shape.leaf_index = i;
+            plan.input_shapes.push_back(shape);
+        }
+
+        std::unordered_map<const ggml_tensor*, int> producer_index;
+        producer_index.reserve(static_cast<size_t>(n_nodes));
+        for (int i = 0; i < n_nodes; ++i) {
+            producer_index[ggml_graph_node(gf, i)] = i;
+        }
+
+        std::vector<Segment> grouped_segments;
+        std::unordered_map<std::string, size_t> group_to_segment;
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            if (!is_graph_cut_tensor(node)) {
+                continue;
+            }
+
+            plan.has_cuts = true;
+            std::string full_name(node->name);
+            std::string payload = full_name.substr(std::strlen(GGML_RUNNER_CUT_PREFIX));
+            size_t sep          = payload.find('|');
+            std::string group   = sep == std::string::npos ? payload : payload.substr(0, sep);
+
+            auto it = group_to_segment.find(group);
+            if (it == group_to_segment.end()) {
+                Segment segment;
+                segment.group_name = group;
+                segment.output_node_indices.push_back(i);
+                group_to_segment[group] = grouped_segments.size();
+                grouped_segments.push_back(std::move(segment));
+            } else {
+                auto& segment = grouped_segments[it->second];
+                segment.output_node_indices.push_back(i);
+            }
+        }
+
+        if (!plan.has_cuts) {
+            return plan;
+        }
+
+        std::unordered_set<int> available_cut_output_node_indices;
+        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
+        for (auto& segment : grouped_segments) {
+            build_segment(gf,
+                          plan,
+                          segment,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          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()) {
+            Segment final_segment;
+            final_segment.group_name = "ggml_runner.final";
+            final_segment.output_node_indices.push_back(n_nodes - 1);
+            build_segment(gf,
+                          plan,
+                          final_segment,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+        }
+
+        return plan;
+    }
+
+    Plan apply_max_vram_budget(ggml_cgraph* gf,
+                               const Plan& base_plan,
+                               size_t max_graph_vram_bytes,
+                               ggml_backend_t backend,
+                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                               const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        int64_t t_budget_begin = ggml_time_ms();
+        if (max_graph_vram_bytes == 0 || !base_plan.has_cuts || base_plan.segments.size() <= 1) {
+            return base_plan;
+        }
+
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        std::unordered_map<const ggml_tensor*, int> producer_index;
+        producer_index.reserve(static_cast<size_t>(n_nodes));
+        for (int i = 0; i < n_nodes; ++i) {
+            producer_index[ggml_graph_node(gf, i)] = i;
+        }
+
+        Plan merged_plan;
+        merged_plan.available = true;
+        merged_plan.has_cuts  = base_plan.has_cuts;
+        merged_plan.valid     = base_plan.valid;
+        merged_plan.n_nodes   = base_plan.n_nodes;
+        merged_plan.n_leafs   = base_plan.n_leafs;
+
+        std::unordered_set<int> available_cut_output_node_indices;
+        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
+
+        size_t start_segment_index = 0;
+        while (start_segment_index < base_plan.segments.size()) {
+            Plan single_plan;
+            auto single_available_cut_output_node_indices = available_cut_output_node_indices;
+            auto single_seed                              = make_segment_seed(base_plan,
+                                                                              start_segment_index,
+                                                                              start_segment_index);
+            build_segment(gf,
+                          single_plan,
+                          single_seed,
+                          producer_index,
+                          single_available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+            GGML_ASSERT(!single_plan.segments.empty());
+
+            size_t best_end_segment_index = start_segment_index;
+            bool can_merge_next_segment   = graph_cut_segment_vram_bytes(single_plan.segments.back()) <= max_graph_vram_bytes;
+
+            while (can_merge_next_segment && best_end_segment_index + 1 < base_plan.segments.size()) {
+                const size_t next_end_segment_index = best_end_segment_index + 1;
+                Plan candidate_plan;
+                auto candidate_available_cut_output_node_indices = available_cut_output_node_indices;
+                auto candidate_seed                              = make_segment_seed(base_plan,
+                                                                                     start_segment_index,
+                                                                                     next_end_segment_index);
+                build_segment(gf,
+                              candidate_plan,
+                              candidate_seed,
+                              producer_index,
+                              candidate_available_cut_output_node_indices,
+                              backend,
+                              params_tensor_set,
+                              log_desc);
+                GGML_ASSERT(!candidate_plan.segments.empty());
+
+                const auto& candidate_segment = candidate_plan.segments.back();
+                if (graph_cut_segment_vram_bytes(candidate_segment) > max_graph_vram_bytes) {
+                    break;
+                }
+
+                best_end_segment_index = next_end_segment_index;
+            }
+
+            auto best_seed = make_segment_seed(base_plan,
+                                               start_segment_index,
+                                               best_end_segment_index);
+            build_segment(gf,
+                          merged_plan,
+                          best_seed,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+            start_segment_index = best_end_segment_index + 1;
+        }
+
+        if (log_desc != nullptr && merged_plan.segments.size() != base_plan.segments.size()) {
+            LOG_INFO("%s graph cut max_vram=%.2f MB merged %zu segments -> %zu segments",
+                     log_desc,
+                     max_graph_vram_bytes / 1024.0 / 1024.0,
+                     base_plan.segments.size(),
+                     merged_plan.segments.size());
+        }
+
+        if (log_desc != nullptr) {
+            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
+                     log_desc,
+                     ggml_time_ms() - t_budget_begin);
+        }
+
+        return merged_plan;
+    }
+
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      size_t max_graph_vram_bytes,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(cache != nullptr);
+
+        int64_t t_prepare_begin = ggml_time_ms();
+        Plan base_plan;
+        int64_t t_plan_begin = ggml_time_ms();
+        if (cache->graph_cut_plan.available && plan_matches_graph(gf, cache->graph_cut_plan)) {
+            base_plan = cache->graph_cut_plan;
+        } else {
+            base_plan                                = build_plan(backend, gf, params_tensor_set, log_desc);
+            cache->graph_cut_plan                    = base_plan;
+            cache->graph_cut_plan.available          = true;
+            cache->budgeted_graph_cut_plan.available = false;
+            if (log_desc != nullptr) {
+                LOG_INFO("%s build cached graph cut plan done (taking %lld ms)", log_desc, ggml_time_ms() - t_plan_begin);
+            }
+        }
+
+        Plan resolved_plan = base_plan;
+        if (max_graph_vram_bytes > 0 && base_plan.has_cuts) {
+            if (cache->budgeted_graph_cut_plan.available &&
+                cache->budgeted_graph_cut_plan_max_vram_bytes == max_graph_vram_bytes &&
+                plan_matches_graph(gf, cache->budgeted_graph_cut_plan)) {
+                resolved_plan = cache->budgeted_graph_cut_plan;
+            } else {
+                resolved_plan                                 = apply_max_vram_budget(gf,
+                                                                                      base_plan,
+                                                                                      max_graph_vram_bytes,
+                                                                                      backend,
+                                                                                      params_tensor_set,
+                                                                                      log_desc);
+                cache->budgeted_graph_cut_plan                = resolved_plan;
+                cache->budgeted_graph_cut_plan.available      = true;
+                cache->budgeted_graph_cut_plan_max_vram_bytes = max_graph_vram_bytes;
+            }
+        }
+        return resolved_plan;
+    }
+
+}  // namespace sd::ggml_graph_cut

src/ggml_graph_cut.h

@@ -0,0 +1,104 @@
+#ifndef __SD_GGML_GRAPH_CUT_H__
+#define __SD_GGML_GRAPH_CUT_H__
+
+#include <array>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+namespace sd::ggml_graph_cut {
+
+    struct Segment {
+        enum InputType {
+            INPUT_EXTERNAL = 0,
+            INPUT_PREVIOUS_CUT,
+            INPUT_PARAM,
+        };
+
+        struct InputRef {
+            InputType type = INPUT_EXTERNAL;
+            std::string display_name;
+            int leaf_index = -1;
+            int node_index = -1;
+        };
+
+        size_t compute_buffer_size      = 0;
+        size_t output_bytes             = 0;
+        size_t input_external_bytes     = 0;
+        size_t input_previous_cut_bytes = 0;
+        size_t input_param_bytes        = 0;
+        std::string group_name;
+        std::vector<int> internal_node_indices;
+        std::vector<int> output_node_indices;
+        std::vector<InputRef> input_refs;
+    };
+
+    struct Plan {
+        struct InputShape {
+            int leaf_index                        = -1;
+            ggml_type type                        = GGML_TYPE_COUNT;
+            std::array<int64_t, GGML_MAX_DIMS> ne = {0, 0, 0, 0};
+        };
+
+        bool available = false;
+        bool has_cuts  = false;
+        bool valid     = true;
+        int n_nodes    = 0;
+        int n_leafs    = 0;
+        std::vector<InputShape> input_shapes;
+        std::vector<Segment> segments;
+    };
+
+    struct PlanCache {
+        Plan graph_cut_plan;
+        Plan budgeted_graph_cut_plan;
+        size_t budgeted_graph_cut_plan_max_vram_bytes = 0;
+    };
+
+    static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:";
+
+    bool is_graph_cut_tensor(const ggml_tensor* tensor);
+    std::string make_graph_cut_name(const std::string& group, const std::string& output);
+    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output);
+    int leaf_count(ggml_cgraph* gf);
+    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index);
+    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor);
+    ggml_tensor* cache_source_tensor(ggml_tensor* tensor);
+    size_t cache_tensor_bytes(const ggml_tensor* tensor);
+    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan);
+    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index);
+    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref);
+    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment);
+    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc);
+    std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
+                                                               const Plan& plan,
+                                                               size_t current_segment_index);
+    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
+                                     const Segment& segment,
+                                     ggml_context** graph_ctx_out);
+    size_t measure_segment_compute_buffer(ggml_backend_t backend,
+                                          ggml_cgraph* gf,
+                                          const Segment& segment,
+                                          const char* log_desc);
+    Plan build_plan(ggml_backend_t backend,
+                    ggml_cgraph* gf,
+                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                    const char* log_desc);
+    Plan apply_max_vram_budget(ggml_cgraph* gf,
+                               const Plan& base_plan,
+                               size_t max_graph_vram_bytes,
+                               ggml_backend_t backend,
+                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                               const char* log_desc);
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      size_t max_graph_vram_bytes,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc);
+}  // namespace sd::ggml_graph_cut
+
+#endif

src/llm.hpp

@@ -346,6 +346,7 @@ namespace LLM {
             auto merger      = std::dynamic_pointer_cast<PatchMerger>(blocks["merger"]);
 
             auto x = patch_embed->forward(ctx, pixel_values);
+            sd::ggml_graph_cut::mark_graph_cut(x, "llm.vision.prelude", "x");
 
             x = ggml_reshape_4d(ctx->ggml_ctx, x, x->ne[0] * spatial_merge_size * spatial_merge_size, x->ne[1] / spatial_merge_size / spatial_merge_size, x->ne[2], x->ne[3]);
             x = ggml_get_rows(ctx->ggml_ctx, x, window_index);
@@ -359,9 +360,11 @@ namespace LLM {
                     mask = nullptr;
                 }
                 x = block->forward(ctx, x, pe, mask);
+                sd::ggml_graph_cut::mark_graph_cut(x, "llm.vision.blocks." + std::to_string(i), "x");
             }
 
             x = merger->forward(ctx, x);
+            sd::ggml_graph_cut::mark_graph_cut(x, "llm.vision.final", "x");
 
             x = ggml_get_rows(ctx->ggml_ctx, x, window_inverse_index);
 
@@ -506,6 +509,7 @@ namespace LLM {
             auto norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
 
             auto x = embed_tokens->forward(ctx, input_ids);
+            sd::ggml_graph_cut::mark_graph_cut(x, "llm.text.prelude", "x");
 
             std::vector<ggml_tensor*> intermediate_outputs;
 
@@ -552,6 +556,10 @@ namespace LLM {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["layers." + std::to_string(i)]);
 
                 x = block->forward(ctx, x, input_pos, attention_mask);
+                if (out_layers.size() > 1) {
+                    x = ggml_cont(ctx->ggml_ctx, x);
+                }
+                sd::ggml_graph_cut::mark_graph_cut(x, "llm.text.layers." + std::to_string(i), "x");
                 if (out_layers.find(i + 1) != out_layers.end()) {
                     intermediate_outputs.push_back(x);
                 }

src/lora.hpp

@@ -129,7 +129,7 @@ struct LoraModel : public GGMLRunner {
         }
     }
 
-    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -152,17 +152,17 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lora_up_name);
             if (iter != lora_tensors.end()) {
-                lora_up = ggml_ext_cast_f32(ctx, iter->second);
+                lora_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lora_mid_name);
             if (iter != lora_tensors.end()) {
-                lora_mid = ggml_ext_cast_f32(ctx, iter->second);
+                lora_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lora_down_name);
             if (iter != lora_tensors.end()) {
-                lora_down = ggml_ext_cast_f32(ctx, iter->second);
+                lora_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             if (lora_up == nullptr || lora_down == nullptr) {
@@ -208,7 +208,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -225,7 +225,7 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(diff_name);
             if (iter != lora_tensors.end()) {
-                curr_updown = ggml_ext_cast_f32(ctx, iter->second);
+                curr_updown = ggml_ext_cast_f32(ctx, backend, iter->second);
             } else {
                 break;
             }
@@ -248,7 +248,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -276,33 +276,33 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(hada_1_down_name);
             if (iter != lora_tensors.end()) {
-                hada_1_down = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_up_name);
             if (iter != lora_tensors.end()) {
-                hada_1_up = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_1_mid = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
                 hada_1_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_1_up));
             }
 
             iter = lora_tensors.find(hada_2_down_name);
             if (iter != lora_tensors.end()) {
-                hada_2_down = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_up_name);
             if (iter != lora_tensors.end()) {
-                hada_2_up = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_2_mid = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
                 hada_2_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_2_up));
             }
 
@@ -351,7 +351,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -378,24 +378,24 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lokr_w1_name);
             if (iter != lora_tensors.end()) {
-                lokr_w1 = ggml_ext_cast_f32(ctx, iter->second);
+                lokr_w1 = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lokr_w2_name);
             if (iter != lora_tensors.end()) {
-                lokr_w2 = ggml_ext_cast_f32(ctx, iter->second);
+                lokr_w2 = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             int64_t rank = 1;
             if (lokr_w1 == nullptr) {
                 iter = lora_tensors.find(lokr_w1_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_a = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w1_a = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w1_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_b = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w1_b = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 if (lokr_w1_a == nullptr || lokr_w1_b == nullptr) {
@@ -410,12 +410,12 @@ struct LoraModel : public GGMLRunner {
             if (lokr_w2 == nullptr) {
                 iter = lora_tensors.find(lokr_w2_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_a = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w2_a = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w2_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_b = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w2_b = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 if (lokr_w2_a == nullptr || lokr_w2_b == nullptr) {
@@ -468,23 +468,23 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
+    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_backend_t backend, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
         // lora
         ggml_tensor* diff = nullptr;
         if (with_lora_and_lokr) {
-            diff = get_lora_weight_diff(model_tensor_name, ctx);
+            diff = get_lora_weight_diff(model_tensor_name, ctx, backend);
         }
         // diff
         if (diff == nullptr) {
-            diff = get_raw_weight_diff(model_tensor_name, ctx);
+            diff = get_raw_weight_diff(model_tensor_name, ctx, backend);
         }
         // loha
         if (diff == nullptr) {
-            diff = get_loha_weight_diff(model_tensor_name, ctx);
+            diff = get_loha_weight_diff(model_tensor_name, ctx, backend);
         }
         // lokr
         if (diff == nullptr && with_lora_and_lokr) {
-            diff = get_lokr_weight_diff(model_tensor_name, ctx);
+            diff = get_lokr_weight_diff(model_tensor_name, ctx, backend);
         }
         if (diff != nullptr) {
             if (ggml_nelements(diff) < ggml_nelements(model_tensor)) {
@@ -502,6 +502,7 @@ struct LoraModel : public GGMLRunner {
     }
 
     ggml_tensor* get_out_diff(ggml_context* ctx,
+                              ggml_backend_t backend,
                               ggml_tensor* x,
                               WeightAdapter::ForwardParams forward_params,
                               const std::string& model_tensor_name) {
@@ -590,7 +591,7 @@ struct LoraModel : public GGMLRunner {
                 }
                 scale_value *= multiplier;
 
-                auto curr_out_diff = ggml_ext_lokr_forward(ctx, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
+                auto curr_out_diff = ggml_ext_lokr_forward(ctx, backend, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
                 if (out_diff == nullptr) {
                     out_diff = curr_out_diff;
                 } else {
@@ -761,7 +762,7 @@ struct LoraModel : public GGMLRunner {
             ggml_tensor* model_tensor     = it.second;
 
             // lora
-            ggml_tensor* diff = get_weight_diff(model_tensor_name, compute_ctx, model_tensor);
+            ggml_tensor* diff = get_weight_diff(model_tensor_name, runtime_backend, compute_ctx, model_tensor);
             if (diff == nullptr) {
                 continue;
             }
@@ -774,7 +775,7 @@ struct LoraModel : public GGMLRunner {
 
             ggml_tensor* final_tensor;
             if (model_tensor->type != GGML_TYPE_F32 && model_tensor->type != GGML_TYPE_F16) {
-                final_tensor = ggml_ext_cast_f32(compute_ctx, model_tensor);
+                final_tensor = ggml_ext_cast_f32(compute_ctx, runtime_backend, model_tensor);
                 final_tensor = ggml_add_inplace(compute_ctx, final_tensor, diff);
                 final_tensor = ggml_cpy(compute_ctx, final_tensor, model_tensor);
             } else {
@@ -841,34 +842,35 @@ public:
         : lora_models(lora_models) {
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
         for (auto& lora_model : lora_models) {
-            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora_and_lokr);
+            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, backend, ctx, weight, with_lora_and_lokr);
             if (diff == nullptr) {
                 continue;
             }
 
             if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
-                weight = ggml_ext_cast_f32(ctx, weight);
+                weight = ggml_ext_cast_f32(ctx, backend, weight);
             }
             weight = ggml_add(ctx, weight, diff);
         }
         return weight;
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name) override {
-        return patch_weight(ctx, weight, weight_name, true);
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name) override {
+        return patch_weight(ctx, backend, weight, weight_name, true);
     }
 
     ggml_tensor* forward_with_lora(ggml_context* ctx,
+                                   ggml_backend_t backend,
                                    ggml_tensor* x,
                                    ggml_tensor* w,
                                    ggml_tensor* b,
                                    const std::string& prefix,
                                    WeightAdapter::ForwardParams forward_params) override {
-        w = patch_weight(ctx, w, prefix + "weight", false);
+        w = patch_weight(ctx, backend, w, prefix + "weight", false);
         if (b) {
-            b = patch_weight(ctx, b, prefix + "bias", false);
+            b = patch_weight(ctx, backend, b, prefix + "bias", false);
         }
         ggml_tensor* out;
         if (forward_params.op_type == ForwardParams::op_type_t::OP_LINEAR) {
@@ -890,7 +892,7 @@ public:
                                    forward_params.conv2d.scale);
         }
         for (auto& lora_model : lora_models) {
-            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, x, forward_params, prefix + "weight");
+            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, backend, x, forward_params, prefix + "weight");
             if (out_diff == nullptr) {
                 continue;
             }

src/mmdit.hpp

@@ -767,6 +767,8 @@ public:
             auto context_x = block->forward(ctx, context, x, c_mod);
             context        = context_x.first;
             x              = context_x.second;
+            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.joint_blocks." + std::to_string(i), "context");
+            sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.joint_blocks." + std::to_string(i), "x");
         }
 
         x = final_layer->forward(ctx, x, c_mod);  // (N, T, patch_size ** 2 * out_channels)
@@ -809,6 +811,11 @@ public:
 
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
         }
+        sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.prelude", "x");
+        sd::ggml_graph_cut::mark_graph_cut(c, "mmdit.prelude", "c");
+        if (context != nullptr) {
+            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.prelude", "context");
+        }
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 

src/model.cpp

@@ -23,24 +23,11 @@
 
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
-#include "ggml-cpu.h"
 #include "ggml.h"
+#include "ggml_extend_backend.hpp"
 #include "zip.h"
 
 #include "name_conversion.h"
-#include "stable-diffusion.h"
-
-#ifdef SD_USE_METAL
-#include "ggml-metal.h"
-#endif
-
-#ifdef SD_USE_VULKAN
-#include "ggml-vulkan.h"
-#endif
-
-#ifdef SD_USE_OPENCL
-#include "ggml-opencl.h"
-#endif
 
 /*================================================= Preprocess ==================================================*/
 

src/preprocessing.hpp

@@ -24,6 +24,75 @@ static inline void preprocessing_set_4d(sd::Tensor<float>& tensor, float value,
     tensor.values()[static_cast<size_t>(preprocessing_offset_4d(tensor, i0, i1, i2, i3))] = value;
 }
 
+static inline uint8_t preprocessing_float_to_u8(float value) {
+    if (value <= 0.0f) {
+        return 0;
+    }
+    if (value >= 1.0f) {
+        return 255;
+    }
+    return static_cast<uint8_t>(value * 255.0f + 0.5f);
+}
+
+static inline void preprocessing_tensor_frame_to_sd_image(const sd::Tensor<float>& tensor, int frame_index, uint8_t* image_data) {
+    const auto& shape = tensor.shape();
+    GGML_ASSERT(shape.size() == 4 || shape.size() == 5);
+    GGML_ASSERT(image_data != nullptr);
+
+    const int width     = static_cast<int>(shape[0]);
+    const int height    = static_cast<int>(shape[1]);
+    const int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
+    const size_t pixels = static_cast<size_t>(width) * static_cast<size_t>(height);
+    const float* src    = tensor.data();
+
+    if (shape.size() == 4) {
+        GGML_ASSERT(frame_index >= 0 && frame_index < shape[3]);
+        const size_t frame_stride = pixels * static_cast<size_t>(channel);
+        const float* frame_ptr    = src + static_cast<size_t>(frame_index) * frame_stride;
+        if (channel == 3) {
+            const float* c0 = frame_ptr;
+            const float* c1 = frame_ptr + pixels;
+            const float* c2 = frame_ptr + pixels * 2;
+            for (size_t i = 0; i < pixels; ++i) {
+                image_data[i * 3 + 0] = preprocessing_float_to_u8(c0[i]);
+                image_data[i * 3 + 1] = preprocessing_float_to_u8(c1[i]);
+                image_data[i * 3 + 2] = preprocessing_float_to_u8(c2[i]);
+            }
+            return;
+        }
+
+        for (size_t i = 0; i < pixels; ++i) {
+            for (int c = 0; c < channel; ++c) {
+                image_data[i * static_cast<size_t>(channel) + static_cast<size_t>(c)] =
+                    preprocessing_float_to_u8(frame_ptr[i + pixels * static_cast<size_t>(c)]);
+            }
+        }
+        return;
+    }
+
+    GGML_ASSERT(frame_index >= 0 && frame_index < shape[2]);
+    const size_t channel_stride = pixels * static_cast<size_t>(shape[2]);
+    const float* frame_ptr      = src + static_cast<size_t>(frame_index) * pixels;
+    if (channel == 3) {
+        const float* c0 = frame_ptr;
+        const float* c1 = frame_ptr + channel_stride;
+        const float* c2 = frame_ptr + channel_stride * 2;
+        for (size_t i = 0; i < pixels; ++i) {
+            image_data[i * 3 + 0] = preprocessing_float_to_u8(c0[i]);
+            image_data[i * 3 + 1] = preprocessing_float_to_u8(c1[i]);
+            image_data[i * 3 + 2] = preprocessing_float_to_u8(c2[i]);
+        }
+        return;
+    }
+
+    for (size_t i = 0; i < pixels; ++i) {
+        for (int c = 0; c < channel; ++c) {
+            image_data[i * static_cast<size_t>(channel) + static_cast<size_t>(c)] =
+                preprocessing_float_to_u8(frame_ptr[i + channel_stride * static_cast<size_t>(c)]);
+        }
+    }
+}
+
 static inline sd::Tensor<float> sd_image_to_preprocessing_tensor(sd_image_t image) {
     sd::Tensor<float> tensor({static_cast<int64_t>(image.width), static_cast<int64_t>(image.height), static_cast<int64_t>(image.channel), 1});
     for (uint32_t y = 0; y < image.height; ++y) {
@@ -39,20 +108,7 @@ static inline sd::Tensor<float> sd_image_to_preprocessing_tensor(sd_image_t imag
 static inline void preprocessing_tensor_to_sd_image(const sd::Tensor<float>& tensor, uint8_t* image_data) {
     GGML_ASSERT(tensor.dim() == 4);
     GGML_ASSERT(tensor.shape()[3] == 1);
-    GGML_ASSERT(image_data != nullptr);
-
-    int width   = static_cast<int>(tensor.shape()[0]);
-    int height  = static_cast<int>(tensor.shape()[1]);
-    int channel = static_cast<int>(tensor.shape()[2]);
-    for (int y = 0; y < height; ++y) {
-        for (int x = 0; x < width; ++x) {
-            for (int c = 0; c < channel; ++c) {
-                float value                               = preprocessing_get_4d(tensor, x, y, c, 0);
-                value                                     = std::min(1.0f, std::max(0.0f, value));
-                image_data[(y * width + x) * channel + c] = static_cast<uint8_t>(std::round(value * 255.0f));
-            }
-        }
-    }
+    preprocessing_tensor_frame_to_sd_image(tensor, 0, image_data);
 }
 
 static inline sd::Tensor<float> gaussian_kernel_tensor(int kernel_size) {

src/qwen_image.hpp

@@ -95,9 +95,7 @@ namespace Qwen {
 
             float scale         = 1.f / 32.f;
             bool force_prec_f32 = false;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
+
             // The purpose of the scale here is to prevent NaN issues in certain situations.
             // For example when using CUDA but the weights are k-quants (not all prompts).
             blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias, false, force_prec_f32, scale));
@@ -124,6 +122,10 @@ namespace Qwen {
             auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
             auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
 
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
             auto norm_added_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_q"]);
             auto norm_added_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_k"]);
 
@@ -410,6 +412,9 @@ namespace Qwen {
             auto img = img_in->forward(ctx, x);
             auto txt = txt_norm->forward(ctx, context);
             txt      = txt_in->forward(ctx, txt);
+            sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.prelude", "txt");
+            // sd::ggml_graph_cut::mark_graph_cut(t_emb, "qwen_image.prelude", "t_emb");
 
             for (int i = 0; i < params.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
@@ -417,6 +422,8 @@ namespace Qwen {
                 auto result = block->forward(ctx, img, txt, t_emb, pe, modulate_index);
                 img         = result.first;
                 txt         = result.second;
+                sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.transformer_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.transformer_blocks." + std::to_string(i), "txt");
             }
 
             if (params.zero_cond_t) {

src/stable-diffusion.cpp

@@ -17,6 +17,7 @@
 #include "pmid.hpp"
 #include "sample-cache.h"
 #include "tae.hpp"
+#include "upscaler.h"
 #include "vae.hpp"
 
 #include "latent-preview.h"
@@ -143,6 +144,7 @@ public:
     std::string taesd_path;
     sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0, 0};
     bool offload_params_to_cpu           = false;
+    float max_vram                       = 0.f;
     bool use_pmid                        = false;
 
     bool is_using_v_parameterization     = false;
@@ -171,60 +173,7 @@ public:
     }
 
     void init_backend() {
-#ifdef SD_USE_CUDA
-        LOG_DEBUG("Using CUDA backend");
-        backend = ggml_backend_cuda_init(0);
-#endif
-#ifdef SD_USE_METAL
-        LOG_DEBUG("Using Metal backend");
-        backend = ggml_backend_metal_init();
-#endif
-#ifdef SD_USE_VULKAN
-        LOG_DEBUG("Using Vulkan backend");
-        size_t device          = 0;
-        const int device_count = ggml_backend_vk_get_device_count();
-        if (device_count) {
-            const char* SD_VK_DEVICE = getenv("SD_VK_DEVICE");
-            if (SD_VK_DEVICE != nullptr) {
-                std::string sd_vk_device_str = SD_VK_DEVICE;
-                try {
-                    device = std::stoull(sd_vk_device_str);
-                } catch (const std::invalid_argument&) {
-                    LOG_WARN("SD_VK_DEVICE environment variable is not a valid integer (%s). Falling back to device 0.", SD_VK_DEVICE);
-                    device = 0;
-                } catch (const std::out_of_range&) {
-                    LOG_WARN("SD_VK_DEVICE environment variable value is out of range for `unsigned long long` type (%s). Falling back to device 0.", SD_VK_DEVICE);
-                    device = 0;
-                }
-                if (device >= device_count) {
-                    LOG_WARN("Cannot find targeted vulkan device (%llu). Falling back to device 0.", device);
-                    device = 0;
-                }
-            }
-            LOG_INFO("Vulkan: Using device %llu", device);
-            backend = ggml_backend_vk_init(device);
-        }
-        if (!backend) {
-            LOG_WARN("Failed to initialize Vulkan backend");
-        }
-#endif
-#ifdef SD_USE_OPENCL
-        LOG_DEBUG("Using OpenCL backend");
-        // ggml_log_set(ggml_log_callback_default, nullptr); // Optional ggml logs
-        backend = ggml_backend_opencl_init();
-        if (!backend) {
-            LOG_WARN("Failed to initialize OpenCL backend");
-        }
-#endif
-#ifdef SD_USE_SYCL
-        LOG_DEBUG("Using SYCL backend");
-        backend = ggml_backend_sycl_init(0);
-#endif
-
-        if (!backend) {
-            LOG_DEBUG("Using CPU backend");
-            backend = ggml_backend_cpu_init();
-        }
+        backend = sd_get_default_backend();
     }
 
     std::shared_ptr<RNG> get_rng(rng_type_t rng_type) {
@@ -242,6 +191,7 @@ public:
         vae_decode_only         = sd_ctx_params->vae_decode_only;
         free_params_immediately = sd_ctx_params->free_params_immediately;
         offload_params_to_cpu   = sd_ctx_params->offload_params_to_cpu;
+        max_vram                = sd_ctx_params->max_vram;
 
         bool use_tae = false;
 
@@ -427,6 +377,10 @@ 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);
+
         {
             clip_backend = backend;
             if (clip_on_cpu && !ggml_backend_is_cpu(backend)) {
@@ -516,6 +470,7 @@ public:
                     clip_vision = std::make_shared<FrozenCLIPVisionEmbedder>(backend,
                                                                              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);
                 }
@@ -592,9 +547,11 @@ 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);
 
+            diffusion_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
             diffusion_model->alloc_params_buffer();
             diffusion_model->get_param_tensors(tensors);
 
@@ -603,6 +560,7 @@ 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);
             }
@@ -675,16 +633,19 @@ public:
             } else if (use_tae && !tae_preview_only) {
                 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");
             } 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");
                 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");
                 }
@@ -1156,9 +1117,14 @@ public:
                     cond_stage_lora_models.push_back(lora);
                 }
             }
+            // Only attach the adapter when there are LoRAs targeting the cond_stage model.
+            // An empty MultiLoraAdapter still routes every linear/conv through
+            // forward_with_lora() instead of the direct kernel path — slower for no benefit.
+            if (!cond_stage_lora_models.empty()) {
                 auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(cond_stage_lora_models);
                 cond_stage_model->set_weight_adapter(multi_lora_adapter);
             }
+        }
         if (diffusion_model) {
             std::vector<std::shared_ptr<LoraModel>> lora_models;
             auto lora_state_diff = lora_state;
@@ -1188,12 +1154,14 @@ public:
                     diffusion_lora_models.push_back(lora);
                 }
             }
+            if (!diffusion_lora_models.empty()) {
                 auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(diffusion_lora_models);
                 diffusion_model->set_weight_adapter(multi_lora_adapter);
                 if (high_noise_diffusion_model) {
                     high_noise_diffusion_model->set_weight_adapter(multi_lora_adapter);
                 }
             }
+        }
 
         if (first_stage_model) {
             std::vector<std::shared_ptr<LoraModel>> lora_models;
@@ -1224,10 +1192,12 @@ public:
                     first_stage_lora_models.push_back(lora);
                 }
             }
+            if (!first_stage_lora_models.empty()) {
                 auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(first_stage_lora_models);
                 first_stage_model->set_weight_adapter(multi_lora_adapter);
             }
         }
+    }
 
     void lora_stat() {
         if (!cond_stage_lora_models.empty()) {
@@ -2113,6 +2083,35 @@ enum lora_apply_mode_t str_to_lora_apply_mode(const char* str) {
     return LORA_APPLY_MODE_COUNT;
 }
 
+const char* hires_upscaler_to_str[] = {
+    "None",
+    "Latent",
+    "Latent (nearest)",
+    "Latent (nearest-exact)",
+    "Latent (antialiased)",
+    "Latent (bicubic)",
+    "Latent (bicubic antialiased)",
+    "Lanczos",
+    "Nearest",
+    "Model",
+};
+
+const char* sd_hires_upscaler_name(enum sd_hires_upscaler_t upscaler) {
+    if (upscaler >= SD_HIRES_UPSCALER_NONE && upscaler < SD_HIRES_UPSCALER_COUNT) {
+        return hires_upscaler_to_str[upscaler];
+    }
+    return NONE_STR;
+}
+
+enum sd_hires_upscaler_t str_to_sd_hires_upscaler(const char* str) {
+    for (int i = 0; i < SD_HIRES_UPSCALER_COUNT; i++) {
+        if (!strcmp(str, hires_upscaler_to_str[i])) {
+            return (enum sd_hires_upscaler_t)i;
+        }
+    }
+    return SD_HIRES_UPSCALER_COUNT;
+}
+
 void sd_cache_params_init(sd_cache_params_t* cache_params) {
     *cache_params                             = {};
     cache_params->mode                        = SD_CACHE_DISABLED;
@@ -2141,6 +2140,19 @@ void sd_cache_params_init(sd_cache_params_t* cache_params) {
     cache_params->spectrum_stop_percent       = 0.9f;
 }
 
+void sd_hires_params_init(sd_hires_params_t* hires_params) {
+    *hires_params                    = {};
+    hires_params->enabled            = false;
+    hires_params->upscaler           = SD_HIRES_UPSCALER_LATENT;
+    hires_params->model_path         = nullptr;
+    hires_params->scale              = 2.0f;
+    hires_params->target_width       = 0;
+    hires_params->target_height      = 0;
+    hires_params->steps              = 0;
+    hires_params->denoising_strength = 0.7f;
+    hires_params->upscale_tile_size  = 128;
+}
+
 void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     *sd_ctx_params                         = {};
     sd_ctx_params->vae_decode_only         = true;
@@ -2152,6 +2164,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->prediction              = PREDICTION_COUNT;
     sd_ctx_params->lora_apply_mode         = LORA_APPLY_AUTO;
     sd_ctx_params->offload_params_to_cpu   = false;
+    sd_ctx_params->max_vram                = 0.f;
     sd_ctx_params->enable_mmap             = false;
     sd_ctx_params->keep_clip_on_cpu        = false;
     sd_ctx_params->keep_control_net_on_cpu = false;
@@ -2193,6 +2206,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "sampler_rng_type: %s\n"
              "prediction: %s\n"
              "offload_params_to_cpu: %s\n"
+             "max_vram: %.3f\n"
              "keep_clip_on_cpu: %s\n"
              "keep_control_net_on_cpu: %s\n"
              "keep_vae_on_cpu: %s\n"
@@ -2225,6 +2239,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              sd_rng_type_name(sd_ctx_params->sampler_rng_type),
              sd_prediction_name(sd_ctx_params->prediction),
              BOOL_STR(sd_ctx_params->offload_params_to_cpu),
+             sd_ctx_params->max_vram,
              BOOL_STR(sd_ctx_params->keep_clip_on_cpu),
              BOOL_STR(sd_ctx_params->keep_control_net_on_cpu),
              BOOL_STR(sd_ctx_params->keep_vae_on_cpu),
@@ -2310,6 +2325,7 @@ void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params) {
     sd_img_gen_params->pm_params         = {nullptr, 0, nullptr, 20.f};
     sd_img_gen_params->vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
     sd_cache_params_init(&sd_img_gen_params->cache);
+    sd_hires_params_init(&sd_img_gen_params->hires);
 }
 
 char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
@@ -2336,7 +2352,8 @@ char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
              "increase_ref_index: %s\n"
              "control_strength: %.2f\n"
              "photo maker: {style_strength = %.2f, id_images_count = %d, id_embed_path = %s}\n"
-             "VAE tiling: %s\n",
+             "VAE tiling: %s\n"
+             "hires: {enabled=%s, upscaler=%s, model_path=%s, scale=%.2f, target=%dx%d, steps=%d, denoising_strength=%.2f}\n",
              SAFE_STR(sd_img_gen_params->prompt),
              SAFE_STR(sd_img_gen_params->negative_prompt),
              sd_img_gen_params->clip_skip,
@@ -2353,7 +2370,15 @@ char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params) {
              sd_img_gen_params->pm_params.style_strength,
              sd_img_gen_params->pm_params.id_images_count,
              SAFE_STR(sd_img_gen_params->pm_params.id_embed_path),
-             BOOL_STR(sd_img_gen_params->vae_tiling_params.enabled));
+             BOOL_STR(sd_img_gen_params->vae_tiling_params.enabled),
+             BOOL_STR(sd_img_gen_params->hires.enabled),
+             sd_hires_upscaler_name(sd_img_gen_params->hires.upscaler),
+             SAFE_STR(sd_img_gen_params->hires.model_path),
+             sd_img_gen_params->hires.scale,
+             sd_img_gen_params->hires.target_width,
+             sd_img_gen_params->hires.target_height,
+             sd_img_gen_params->hires.steps,
+             sd_img_gen_params->hires.denoising_strength);
     const char* cache_mode_str = "disabled";
     if (sd_img_gen_params->cache.mode == SD_CACHE_EASYCACHE) {
         cache_mode_str = "easycache";
@@ -2534,6 +2559,7 @@ struct GenerationRequest {
     sd_guidance_params_t guidance            = {};
     sd_guidance_params_t high_noise_guidance = {};
     sd_pm_params_t pm_params                 = {};
+    sd_hires_params_t hires                  = {};
     int frames                               = -1;
     float vace_strength                      = 1.f;
 
@@ -2555,6 +2581,7 @@ struct GenerationRequest {
         auto_resize_ref_image       = sd_img_gen_params->auto_resize_ref_image;
         guidance                    = sd_img_gen_params->sample_params.guidance;
         pm_params                   = sd_img_gen_params->pm_params;
+        hires                       = sd_img_gen_params->hires;
         cache_params                = &sd_img_gen_params->cache;
         resolve(sd_ctx);
     }
@@ -2577,26 +2604,76 @@ struct GenerationRequest {
     }
 
     void align_generation_request_size() {
+        align_image_size(&width, &height, "generation request");
+    }
+
+    void align_image_size(int* target_width, int* target_height, const char* label) {
         int spatial_multiple = vae_scale_factor * diffusion_model_down_factor;
-        int width_offset     = align_up_offset(width, spatial_multiple);
-        int height_offset    = align_up_offset(height, spatial_multiple);
+        int width_offset     = align_up_offset(*target_width, spatial_multiple);
+        int height_offset    = align_up_offset(*target_height, spatial_multiple);
         if (width_offset <= 0 && height_offset <= 0) {
             return;
         }
 
-        int original_width  = width;
-        int original_height = height;
+        int original_width  = *target_width;
+        int original_height = *target_height;
 
-        width += width_offset;
-        height += height_offset;
-        LOG_WARN("align up %dx%d to %dx%d (multiple=%d)",
+        *target_width += width_offset;
+        *target_height += height_offset;
+        LOG_WARN("align %s up %dx%d to %dx%d (multiple=%d)",
+                 label,
                  original_width,
                  original_height,
-                 width,
-                 height,
+                 *target_width,
+                 *target_height,
                  spatial_multiple);
     }
 
+    void resolve_hires() {
+        if (!hires.enabled) {
+            return;
+        }
+        if (hires.upscaler == SD_HIRES_UPSCALER_NONE) {
+            hires.enabled = false;
+            return;
+        }
+        if (hires.upscaler < SD_HIRES_UPSCALER_NONE || hires.upscaler >= SD_HIRES_UPSCALER_COUNT) {
+            LOG_WARN("hires upscaler '%d' is invalid, disabling hires", hires.upscaler);
+            hires.enabled = false;
+            return;
+        }
+        if (hires.upscaler == SD_HIRES_UPSCALER_MODEL && strlen(SAFE_STR(hires.model_path)) == 0) {
+            LOG_WARN("hires model upscaler requires a model path, disabling hires");
+            hires.enabled = false;
+            return;
+        }
+        if (hires.scale <= 0.f && hires.target_width <= 0 && hires.target_height <= 0) {
+            LOG_WARN("hires scale must be positive when no target size is set, disabling hires");
+            hires.enabled = false;
+            return;
+        }
+        hires.denoising_strength = std::clamp(hires.denoising_strength, 0.0001f, 1.f);
+        hires.steps              = std::max(0, hires.steps);
+
+        if (hires.target_width > 0 && hires.target_height > 0) {
+            // pass
+        } else if (hires.target_width > 0) {
+            hires.target_height = hires.target_width;
+        } else if (hires.target_height > 0) {
+            hires.target_width = hires.target_height;
+        } else {
+            hires.target_width  = static_cast<int>(std::round(width * hires.scale));
+            hires.target_height = static_cast<int>(std::round(height * hires.scale));
+        }
+
+        if (hires.target_width <= 0 || hires.target_height <= 0) {
+            LOG_WARN("hires target size is not positive, disabling hires");
+            hires.enabled = false;
+            return;
+        }
+        align_image_size(&hires.target_width, &hires.target_height, "hires target");
+    }
+
     static void resolve_guidance(sd_ctx_t* sd_ctx,
                                  sd_guidance_params_t* guidance,
                                  bool* use_uncond,
@@ -2637,6 +2714,7 @@ struct GenerationRequest {
 
     void resolve(sd_ctx_t* sd_ctx) {
         align_generation_request_size();
+        resolve_hires();
         seed = resolve_seed(seed);
 
         resolve_guidance(sd_ctx, &guidance, &use_uncond, &use_img_cond);
@@ -3127,7 +3205,7 @@ static sd_image_t* decode_image_outputs(sd_ctx_t* sd_ctx,
         }
         decoded_images.push_back(std::move(image));
         int64_t t2 = ggml_time_ms();
-        LOG_INFO("latent %" PRId64 " decoded, taking %.2fs", i + 1, (t2 - t1) * 1.0f / 1000);
+        LOG_INFO("latent %zu decoded, taking %.2fs", i + 1, (t2 - t1) * 1.0f / 1000);
     }
 
     int64_t t4 = ggml_time_ms();
@@ -3149,6 +3227,135 @@ static sd_image_t* decode_image_outputs(sd_ctx_t* sd_ctx,
     return result_images;
 }
 
+static sd::Tensor<float> upscale_hires_latent(sd_ctx_t* sd_ctx,
+                                              const sd::Tensor<float>& latent,
+                                              const GenerationRequest& request,
+                                              UpscalerGGML* upscaler) {
+    auto get_hires_latent_target_shape = [&]() {
+        std::vector<int64_t> target_shape = latent.shape();
+        if (target_shape.size() < 2) {
+            target_shape.clear();
+            return target_shape;
+        }
+        target_shape[0] = request.hires.target_width / request.vae_scale_factor;
+        target_shape[1] = request.hires.target_height / request.vae_scale_factor;
+        return target_shape;
+    };
+
+    if (request.hires.upscaler == SD_HIRES_UPSCALER_LATENT ||
+        request.hires.upscaler == SD_HIRES_UPSCALER_LATENT_NEAREST ||
+        request.hires.upscaler == SD_HIRES_UPSCALER_LATENT_NEAREST_EXACT ||
+        request.hires.upscaler == SD_HIRES_UPSCALER_LATENT_ANTIALIASED ||
+        request.hires.upscaler == SD_HIRES_UPSCALER_LATENT_BICUBIC ||
+        request.hires.upscaler == SD_HIRES_UPSCALER_LATENT_BICUBIC_ANTIALIASED) {
+        std::vector<int64_t> target_shape = get_hires_latent_target_shape();
+        if (target_shape.empty()) {
+            LOG_ERROR("latent has invalid shape for hires upscale");
+            return {};
+        }
+
+        sd::ops::InterpolateMode mode = sd::ops::InterpolateMode::Nearest;
+        bool antialias                = false;
+        switch (request.hires.upscaler) {
+            case SD_HIRES_UPSCALER_LATENT:
+                mode = sd::ops::InterpolateMode::Bilinear;
+                break;
+            case SD_HIRES_UPSCALER_LATENT_NEAREST:
+                mode = sd::ops::InterpolateMode::Nearest;
+                break;
+            case SD_HIRES_UPSCALER_LATENT_NEAREST_EXACT:
+                mode = sd::ops::InterpolateMode::NearestExact;
+                break;
+            case SD_HIRES_UPSCALER_LATENT_ANTIALIASED:
+                mode      = sd::ops::InterpolateMode::Bilinear;
+                antialias = true;
+                break;
+            case SD_HIRES_UPSCALER_LATENT_BICUBIC:
+                mode = sd::ops::InterpolateMode::Bicubic;
+                break;
+            case SD_HIRES_UPSCALER_LATENT_BICUBIC_ANTIALIASED:
+                mode      = sd::ops::InterpolateMode::Bicubic;
+                antialias = true;
+                break;
+            default:
+                break;
+        }
+
+        LOG_INFO("hires %s upscale %" PRId64 "x%" PRId64 " -> %" PRId64 "x%" PRId64,
+                 sd_hires_upscaler_name(request.hires.upscaler),
+                 latent.shape()[0],
+                 latent.shape()[1],
+                 target_shape[0],
+                 target_shape[1]);
+
+        return sd::ops::interpolate(latent, target_shape, mode, false, antialias);
+    } else if (request.hires.upscaler == SD_HIRES_UPSCALER_MODEL ||
+               request.hires.upscaler == SD_HIRES_UPSCALER_LANCZOS ||
+               request.hires.upscaler == SD_HIRES_UPSCALER_NEAREST) {
+        if (sd_ctx->sd->vae_decode_only) {
+            LOG_ERROR("hires %s upscaler requires VAE encoder weights; create the context with vae_decode_only=false",
+                      sd_hires_upscaler_name(request.hires.upscaler));
+            return {};
+        }
+        if (request.hires.upscaler == SD_HIRES_UPSCALER_MODEL && upscaler == nullptr) {
+            LOG_ERROR("hires model upscaler context is null");
+            return {};
+        }
+
+        sd::Tensor<float> decoded = sd_ctx->sd->decode_first_stage(latent);
+        if (decoded.empty()) {
+            LOG_ERROR("decode_first_stage failed before hires %s upscale",
+                      sd_hires_upscaler_name(request.hires.upscaler));
+            return {};
+        }
+
+        sd::Tensor<float> upscaled_tensor;
+        if (request.hires.upscaler == SD_HIRES_UPSCALER_MODEL) {
+            upscaled_tensor = upscaler->upscale_tensor(decoded);
+            if (upscaled_tensor.empty()) {
+                LOG_ERROR("hires model upscale failed");
+                return {};
+            }
+
+            if (upscaled_tensor.shape()[0] != request.hires.target_width ||
+                upscaled_tensor.shape()[1] != request.hires.target_height) {
+                upscaled_tensor = sd::ops::interpolate(upscaled_tensor,
+                                                       {request.hires.target_width,
+                                                        request.hires.target_height,
+                                                        upscaled_tensor.shape()[2],
+                                                        upscaled_tensor.shape()[3]});
+            }
+        } else {
+            sd::ops::InterpolateMode mode = request.hires.upscaler == SD_HIRES_UPSCALER_LANCZOS
+                                                ? sd::ops::InterpolateMode::Lanczos
+                                                : sd::ops::InterpolateMode::Nearest;
+            LOG_INFO("hires %s image upscale %" PRId64 "x%" PRId64 " -> %dx%d",
+                     sd_hires_upscaler_name(request.hires.upscaler),
+                     decoded.shape()[0],
+                     decoded.shape()[1],
+                     request.hires.target_width,
+                     request.hires.target_height);
+            upscaled_tensor = sd::ops::interpolate(decoded,
+                                                   {request.hires.target_width,
+                                                    request.hires.target_height,
+                                                    decoded.shape()[2],
+                                                    decoded.shape()[3]},
+                                                   mode);
+            upscaled_tensor = sd::ops::clamp(upscaled_tensor, 0.0f, 1.0f);
+        }
+
+        sd::Tensor<float> upscaled_latent = sd_ctx->sd->encode_first_stage(upscaled_tensor);
+        if (upscaled_latent.empty()) {
+            LOG_ERROR("encode_first_stage failed after hires %s upscale",
+                      sd_hires_upscaler_name(request.hires.upscaler));
+        }
+        return upscaled_latent;
+    }
+
+    LOG_ERROR("unsupported hires upscaler '%s'", sd_hires_upscaler_name(request.hires.upscaler));
+    return {};
+}
+
 SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params) {
     if (sd_ctx == nullptr || sd_img_gen_params == nullptr) {
         return nullptr;
@@ -3236,14 +3443,143 @@ SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* s
         }
         return nullptr;
     }
-    if (sd_ctx->sd->free_params_immediately) {
+    if (sd_ctx->sd->free_params_immediately && !request.hires.enabled) {
         sd_ctx->sd->diffusion_model->free_params_buffer();
     }
     int64_t denoise_end = ggml_time_ms();
-    LOG_INFO("generating %" PRId64 " latent images completed, taking %.2fs",
+    LOG_INFO("generating %zu latent images completed, taking %.2fs",
              final_latents.size(),
              (denoise_end - denoise_start) * 1.0f / 1000);
 
+    if (request.hires.enabled && request.hires.target_width > 0) {
+        LOG_INFO("hires fix: upscaling to %dx%d", request.hires.target_width, request.hires.target_height);
+
+        std::unique_ptr<UpscalerGGML> hires_upscaler;
+        if (request.hires.upscaler == SD_HIRES_UPSCALER_MODEL) {
+            LOG_INFO("hires fix: loading model upscaler from '%s'", request.hires.model_path);
+            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);
+            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,
+                                                sd_ctx->sd->n_threads)) {
+                LOG_ERROR("load hires model upscaler failed");
+                if (sd_ctx->sd->free_params_immediately) {
+                    sd_ctx->sd->diffusion_model->free_params_buffer();
+                }
+                return nullptr;
+            }
+        }
+
+        int hires_steps = request.hires.steps > 0 ? request.hires.steps : plan.sample_steps;
+
+        // sd-webui behavior: scale up total steps so trimming by denoising_strength yields exactly hires_steps effective steps,
+        // unlike img2img which trims from a fixed step count
+        hires_steps = static_cast<int>(hires_steps / request.hires.denoising_strength);
+
+        std::vector<float> hires_sigmas = sd_ctx->sd->denoiser->get_sigmas(
+            hires_steps,
+            sd_ctx->sd->get_image_seq_len(request.hires.target_height, request.hires.target_width),
+            sd_img_gen_params->sample_params.scheduler,
+            sd_ctx->sd->version);
+
+        size_t t_enc = static_cast<size_t>(hires_steps * request.hires.denoising_strength);
+        if (t_enc >= static_cast<size_t>(hires_steps)) {
+            t_enc = static_cast<size_t>(hires_steps) - 1;
+        }
+        std::vector<float> hires_sigma_sched(hires_sigmas.begin() + hires_steps - static_cast<int>(t_enc) - 1,
+                                             hires_sigmas.end());
+        LOG_INFO("hires fix: %d steps, denoising_strength=%.2f, sigma_sched_size=%zu",
+                 hires_steps,
+                 request.hires.denoising_strength,
+                 hires_sigma_sched.size());
+
+        std::vector<sd::Tensor<float>> hires_final_latents;
+        int64_t hires_denoise_start = ggml_time_ms();
+        for (int b = 0; b < (int)final_latents.size(); b++) {
+            int64_t cur_seed = request.seed + b;
+            sd_ctx->sd->rng->manual_seed(cur_seed);
+            sd_ctx->sd->sampler_rng->manual_seed(cur_seed);
+
+            sd::Tensor<float> upscaled = upscale_hires_latent(sd_ctx,
+                                                              final_latents[b],
+                                                              request,
+                                                              hires_upscaler.get());
+            if (upscaled.empty()) {
+                if (sd_ctx->sd->free_params_immediately) {
+                    sd_ctx->sd->diffusion_model->free_params_buffer();
+                }
+                return nullptr;
+            }
+
+            sd::Tensor<float> noise = sd::randn_like<float>(upscaled, sd_ctx->sd->rng);
+
+            sd::Tensor<float> hires_denoise_mask;
+            if (!latents.denoise_mask.empty()) {
+                std::vector<int64_t> mask_shape = latents.denoise_mask.shape();
+                mask_shape[0]                   = upscaled.shape()[0];
+                mask_shape[1]                   = upscaled.shape()[1];
+                hires_denoise_mask              = sd::ops::interpolate(latents.denoise_mask,
+                                                                       mask_shape,
+                                                                       sd::ops::InterpolateMode::NearestMax);
+            }
+
+            int64_t hires_sample_start = ggml_time_ms();
+            sd::Tensor<float> x_0      = sd_ctx->sd->sample(sd_ctx->sd->diffusion_model,
+                                                            true,
+                                                            upscaled,
+                                                            std::move(noise),
+                                                            embeds.cond,
+                                                            embeds.uncond,
+                                                            embeds.img_cond,
+                                                            embeds.id_cond,
+                                                            latents.control_image,
+                                                            request.control_strength,
+                                                            request.guidance,
+                                                            plan.eta,
+                                                            request.shifted_timestep,
+                                                            plan.sample_method,
+                                                            sd_ctx->sd->is_flow_denoiser(),
+                                                            hires_sigma_sched,
+                                                            plan.start_merge_step,
+                                                            latents.ref_latents,
+                                                            request.increase_ref_index,
+                                                            hires_denoise_mask,
+                                                            sd::Tensor<float>(),
+                                                            1.f,
+                                                            request.cache_params);
+            int64_t hires_sample_end   = ggml_time_ms();
+            if (!x_0.empty()) {
+                LOG_INFO("hires sampling %d/%d completed, taking %.2fs",
+                         b + 1,
+                         (int)final_latents.size(),
+                         (hires_sample_end - hires_sample_start) * 1.0f / 1000);
+                hires_final_latents.push_back(std::move(x_0));
+                continue;
+            }
+
+            LOG_ERROR("hires sampling for image %d/%d failed after %.2fs",
+                      b + 1,
+                      (int)final_latents.size(),
+                      (hires_sample_end - hires_sample_start) * 1.0f / 1000);
+            if (sd_ctx->sd->free_params_immediately) {
+                sd_ctx->sd->diffusion_model->free_params_buffer();
+            }
+            return nullptr;
+        }
+        if (sd_ctx->sd->free_params_immediately) {
+            sd_ctx->sd->diffusion_model->free_params_buffer();
+        }
+        int64_t hires_denoise_end = ggml_time_ms();
+        LOG_INFO("hires fix completed, taking %.2fs", (hires_denoise_end - hires_denoise_start) * 1.0f / 1000);
+
+        final_latents = std::move(hires_final_latents);
+    }
+
     auto result = decode_image_outputs(sd_ctx, request, final_latents);
     if (result == nullptr) {
         return nullptr;

src/t5.hpp

@@ -251,7 +251,8 @@ public:
                          ggml_tensor* x,
                          ggml_tensor* past_bias                = nullptr,
                          ggml_tensor* attention_mask           = nullptr,
-                         ggml_tensor* relative_position_bucket = nullptr) {
+                         ggml_tensor* relative_position_bucket = nullptr,
+                         const std::string& graph_cut_prefix   = "") {
         // x: [N, n_token, model_dim]
         for (int i = 0; i < num_layers; i++) {
             auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
@@ -259,6 +260,9 @@ public:
             auto ret  = block->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
             x         = ret.first;
             past_bias = ret.second;
+            if (!graph_cut_prefix.empty()) {
+                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".block." + std::to_string(i), "x");
+            }
         }
 
         auto final_layer_norm = std::dynamic_pointer_cast<T5LayerNorm>(blocks["final_layer_norm"]);
@@ -305,7 +309,8 @@ public:
         auto encoder = std::dynamic_pointer_cast<T5Stack>(blocks["encoder"]);
 
         auto x = shared->forward(ctx, input_ids);
-        x      = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
+        sd::ggml_graph_cut::mark_graph_cut(x, "t5.prelude", "x");
+        x = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket, "t5");
         return x;
     }
 };

src/tensor.hpp

@@ -815,11 +815,202 @@ namespace sd {
     namespace ops {
         enum class InterpolateMode {
             Nearest,
+            NearestExact,
             NearestMax,
             NearestMin,
             NearestAvg,
+            Bilinear,
+            Bicubic,
+            Lanczos,
         };
 
+        inline bool is_nearest_like_interpolate_mode(InterpolateMode mode) {
+            return mode == InterpolateMode::Nearest ||
+                   mode == InterpolateMode::NearestExact ||
+                   mode == InterpolateMode::NearestMax ||
+                   mode == InterpolateMode::NearestMin ||
+                   mode == InterpolateMode::NearestAvg;
+        }
+
+        inline bool is_2d_filter_interpolate_mode(InterpolateMode mode) {
+            return mode == InterpolateMode::Bilinear ||
+                   mode == InterpolateMode::Bicubic ||
+                   mode == InterpolateMode::Lanczos;
+        }
+
+        inline int64_t nearest_exact_interpolate_index(int64_t output_index,
+                                                       int64_t input_size,
+                                                       int64_t output_size) {
+            const double scale  = static_cast<double>(input_size) / static_cast<double>(output_size);
+            const double center = (static_cast<double>(output_index) + 0.5) * scale - 0.5;
+            return std::min(std::max<int64_t>(static_cast<int64_t>(std::floor(center + 0.5)), 0), input_size - 1);
+        }
+
+        inline double linear_interpolate_weight(double x) {
+            x = std::abs(x);
+            return x < 1.0 ? 1.0 - x : 0.0;
+        }
+
+        inline double cubic_interpolate_weight(double x) {
+            constexpr double a = -0.75;  // Match PyTorch bicubic interpolation.
+            x                  = std::abs(x);
+            if (x <= 1.0) {
+                return ((a + 2.0) * x - (a + 3.0)) * x * x + 1.0;
+            }
+            if (x < 2.0) {
+                return ((a * x - 5.0 * a) * x + 8.0 * a) * x - 4.0 * a;
+            }
+            return 0.0;
+        }
+
+        inline double sinc(double x) {
+            constexpr double pi = 3.14159265358979323846;
+            if (std::abs(x) < 1e-12) {
+                return 1.0;
+            }
+            const double pix = pi * x;
+            return std::sin(pix) / pix;
+        }
+
+        inline double lanczos_interpolate_weight(double x) {
+            constexpr double radius = 3.0;
+            x                       = std::abs(x);
+            if (x >= radius) {
+                return 0.0;
+            }
+            return sinc(x) * sinc(x / radius);
+        }
+
+        struct InterpolateContributor {
+            int64_t index;
+            double weight;
+        };
+
+        inline std::vector<std::vector<InterpolateContributor>> make_interpolate_contributors(
+            int64_t input_size,
+            int64_t output_size,
+            InterpolateMode mode,
+            bool antialias) {
+            std::vector<std::vector<InterpolateContributor>> contributors(static_cast<size_t>(output_size));
+            const double scale        = static_cast<double>(input_size) / static_cast<double>(output_size);
+            const double filter_scale = antialias ? std::max(1.0, scale) : 1.0;
+
+            for (int64_t out = 0; out < output_size; ++out) {
+                const double center = (static_cast<double>(out) + 0.5) * scale - 0.5;
+                int64_t start       = 0;
+                int64_t end         = 0;
+
+                if (mode == InterpolateMode::Bilinear) {
+                    const double support = filter_scale;
+                    start                = static_cast<int64_t>(std::ceil(center - support));
+                    end                  = static_cast<int64_t>(std::floor(center + support));
+                } else if (mode == InterpolateMode::Bicubic) {
+                    const double support = 2.0 * filter_scale;
+                    start                = static_cast<int64_t>(std::ceil(center - support));
+                    end                  = static_cast<int64_t>(std::floor(center + support));
+                } else if (mode == InterpolateMode::Lanczos) {
+                    const double support = 3.0 * filter_scale;
+                    start                = static_cast<int64_t>(std::ceil(center - support));
+                    end                  = static_cast<int64_t>(std::floor(center + support));
+                } else {
+                    tensor_throw_invalid_argument("Unsupported 2D filter interpolate mode: mode=" +
+                                                  std::to_string(static_cast<int>(mode)));
+                }
+
+                double weight_sum                                      = 0.0;
+                std::vector<InterpolateContributor>& axis_contributors = contributors[static_cast<size_t>(out)];
+                axis_contributors.reserve(static_cast<size_t>(end - start + 1));
+
+                for (int64_t in = start; in <= end; ++in) {
+                    double weight = 0.0;
+                    if (mode == InterpolateMode::Bilinear) {
+                        weight = linear_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
+                    } else if (mode == InterpolateMode::Bicubic) {
+                        weight = cubic_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
+                    } else {
+                        weight = lanczos_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
+                    }
+
+                    if (weight == 0.0) {
+                        continue;
+                    }
+
+                    const int64_t clamped_index = std::min(std::max<int64_t>(in, 0), input_size - 1);
+                    axis_contributors.push_back({clamped_index, weight});
+                    weight_sum += weight;
+                }
+
+                if ((antialias || mode == InterpolateMode::Lanczos) &&
+                    std::abs(weight_sum) > 1e-12) {
+                    for (auto& contributor : axis_contributors) {
+                        contributor.weight /= weight_sum;
+                    }
+                }
+
+                if (axis_contributors.empty()) {
+                    const int64_t nearest = std::min(
+                        std::max<int64_t>(static_cast<int64_t>(std::floor(center + 0.5)), 0),
+                        input_size - 1);
+                    axis_contributors.push_back({nearest, 1.0});
+                }
+            }
+
+            return contributors;
+        }
+
+        template <typename T>
+        inline Tensor<T> interpolate_2d_filter(const Tensor<T>& input,
+                                               const std::vector<int64_t>& output_shape,
+                                               InterpolateMode mode,
+                                               bool antialias) {
+            if (input.dim() < 2) {
+                tensor_throw_invalid_argument("2D filter interpolate requires rank >= 2: input_shape=" +
+                                              tensor_shape_to_string(input.shape()) + ", output_shape=" +
+                                              tensor_shape_to_string(output_shape));
+            }
+            for (size_t i = 2; i < output_shape.size(); ++i) {
+                if (input.shape()[i] != output_shape[i]) {
+                    tensor_throw_invalid_argument("2D filter interpolate only supports resizing dimensions 0 and 1: input_shape=" +
+                                                  tensor_shape_to_string(input.shape()) + ", output_shape=" +
+                                                  tensor_shape_to_string(output_shape));
+                }
+            }
+
+            Tensor<T> output(output_shape);
+            const int64_t input_width   = input.shape()[0];
+            const int64_t input_height  = input.shape()[1];
+            const int64_t output_width  = output_shape[0];
+            const int64_t output_height = output_shape[1];
+            const int64_t input_plane   = input_width * input_height;
+            const int64_t output_plane  = output_width * output_height;
+            const int64_t plane_count   = input.numel() / input_plane;
+
+            auto x_contributors = make_interpolate_contributors(input_width, output_width, mode, antialias);
+            auto y_contributors = make_interpolate_contributors(input_height, output_height, mode, antialias);
+
+            for (int64_t plane = 0; plane < plane_count; ++plane) {
+                const int64_t input_plane_offset  = plane * input_plane;
+                const int64_t output_plane_offset = plane * output_plane;
+                for (int64_t y = 0; y < output_height; ++y) {
+                    const auto& y_axis = y_contributors[static_cast<size_t>(y)];
+                    for (int64_t x = 0; x < output_width; ++x) {
+                        const auto& x_axis = x_contributors[static_cast<size_t>(x)];
+                        double value       = 0.0;
+                        for (const auto& yc : y_axis) {
+                            const int64_t input_row_offset = input_plane_offset + yc.index * input_width;
+                            for (const auto& xc : x_axis) {
+                                value += static_cast<double>(input.data()[input_row_offset + xc.index]) *
+                                         xc.weight * yc.weight;
+                            }
+                        }
+                        output.data()[output_plane_offset + y * output_width + x] = static_cast<T>(value);
+                    }
+                }
+            }
+
+            return output;
+        }
+
         inline int64_t normalize_slice_bound(int64_t index, int64_t dim_size) {
             if (index < 0) {
                 index += dim_size;
@@ -1014,17 +1205,20 @@ namespace sd {
         inline Tensor<T> interpolate(const Tensor<T>& input,
                                      std::vector<int64_t> output_shape,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false) {
-            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
-                                               mode == InterpolateMode::NearestMax ||
-                                               mode == InterpolateMode::NearestMin ||
-                                               mode == InterpolateMode::NearestAvg);
-            if (!is_nearest_like_mode) {
-                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
+                                     bool align_corners   = false,
+                                     bool antialias       = false) {
+            const bool is_nearest_like_mode = is_nearest_like_interpolate_mode(mode);
+            const bool is_2d_filter_mode    = is_2d_filter_interpolate_mode(mode);
+            if (!is_nearest_like_mode && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+                                              std::to_string(static_cast<int>(mode)));
+            }
+            if (antialias && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Tensor interpolate antialias requires a 2D filter mode: mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for tensor interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()) + ", output_shape=" +
                                               tensor_shape_to_string(output_shape));
             }
@@ -1051,6 +1245,10 @@ namespace sd {
                 }
             }
 
+            if (is_2d_filter_mode) {
+                return interpolate_2d_filter(input, output_shape, mode, antialias);
+            }
+
             bool has_downsampling = false;
             for (int64_t i = 0; i < input.dim(); ++i) {
                 if (input.shape()[i] > output_shape[i]) {
@@ -1060,13 +1258,21 @@ namespace sd {
             }
 
             Tensor<T> output(std::move(output_shape));
-            if (mode == InterpolateMode::Nearest || !has_downsampling) {
+            if (mode == InterpolateMode::Nearest ||
+                mode == InterpolateMode::NearestExact ||
+                !has_downsampling) {
                 for (int64_t flat = 0; flat < output.numel(); ++flat) {
                     std::vector<int64_t> output_coord = tensor_unravel_index(flat, output.shape());
                     std::vector<int64_t> input_coord(static_cast<size_t>(input.dim()), 0);
                     for (size_t i = 0; i < static_cast<size_t>(input.dim()); ++i) {
+                        if (mode == InterpolateMode::NearestExact) {
+                            input_coord[i] = nearest_exact_interpolate_index(output_coord[i],
+                                                                             input.shape()[i],
+                                                                             output.shape()[i]);
+                        } else {
                             input_coord[i] = output_coord[i] * input.shape()[i] / output.shape()[i];
                         }
+                    }
                     output[flat] = input.index(input_coord);
                 }
 
@@ -1083,6 +1289,12 @@ namespace sd {
                         return T(0);
                     case InterpolateMode::Nearest:
                         return T(0);
+                    case InterpolateMode::NearestExact:
+                        return T(0);
+                    case InterpolateMode::Bilinear:
+                    case InterpolateMode::Bicubic:
+                    case InterpolateMode::Lanczos:
+                        break;
                 }
 
                 tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
@@ -1102,6 +1314,12 @@ namespace sd {
                         break;
                     case InterpolateMode::Nearest:
                         break;
+                    case InterpolateMode::NearestExact:
+                        break;
+                    case InterpolateMode::Bilinear:
+                    case InterpolateMode::Bicubic:
+                    case InterpolateMode::Lanczos:
+                        break;
                 }
             };
 
@@ -1157,17 +1375,20 @@ namespace sd {
                                      const std::optional<std::vector<int64_t>>& size,
                                      const std::optional<std::vector<double>>& scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false) {
-            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
-                                               mode == InterpolateMode::NearestMax ||
-                                               mode == InterpolateMode::NearestMin ||
-                                               mode == InterpolateMode::NearestAvg);
-            if (!is_nearest_like_mode) {
-                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
+                                     bool align_corners   = false,
+                                     bool antialias       = false) {
+            const bool is_nearest_like_mode = is_nearest_like_interpolate_mode(mode);
+            const bool is_2d_filter_mode    = is_2d_filter_interpolate_mode(mode);
+            if (!is_nearest_like_mode && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+                                              std::to_string(static_cast<int>(mode)));
+            }
+            if (antialias && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Tensor interpolate antialias requires a 2D filter mode: mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for tensor interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()));
             }
             if (size.has_value() == scale_factor.has_value()) {
@@ -1211,7 +1432,7 @@ namespace sd {
                 }
             }
 
-            return interpolate(input, std::move(output_shape), mode, align_corners);
+            return interpolate(input, std::move(output_shape), mode, align_corners, antialias);
         }
 
         template <typename T>
@@ -1219,12 +1440,14 @@ namespace sd {
                                      const std::optional<std::vector<int64_t>>& size,
                                      double scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false) {
+                                     bool align_corners   = false,
+                                     bool antialias       = false) {
             return interpolate(input,
                                size,
                                std::vector<double>(size.has_value() ? size->size() : input.dim(), scale_factor),
                                mode,
-                               align_corners);
+                               align_corners,
+                               antialias);
         }
 
         template <typename T>

src/tokenizers/clip_tokenizer.cpp

@@ -62,7 +62,7 @@ void CLIPTokenizer::load_from_merges(const std::string& merges_utf8_str) {
     }
     vocab.push_back(utf8_to_utf32("<|startoftext|>"));
     vocab.push_back(utf8_to_utf32("<|endoftext|>"));
-    LOG_DEBUG("vocab size: %llu", vocab.size());
+    LOG_DEBUG("vocab size: %zu", vocab.size());
     int i = 0;
     for (const auto& token : vocab) {
         encoder[token] = i;

src/tokenizers/mistral_tokenizer.cpp

@@ -28,7 +28,7 @@ void MistralTokenizer::load_from_merges(const std::string& merges_utf8_str, cons
         byte_decoder[pair.second] = pair.first;
     }
     std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
-    LOG_DEBUG("merges size %llu", merges.size());
+    LOG_DEBUG("merges size %zu", merges.size());
     std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
     for (const auto& merge : merges) {
         size_t space_pos = merge.find(' ');

src/tokenizers/qwen2_tokenizer.cpp

@@ -11,7 +11,7 @@ void Qwen2Tokenizer::load_from_merges(const std::string& merges_utf8_str) {
     }
 
     std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
-    LOG_DEBUG("merges size %llu", merges.size());
+    LOG_DEBUG("merges size %zu", merges.size());
     std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
     for (const auto& merge : merges) {
         size_t space_pos = merge.find(' ');

src/unet.hpp

@@ -482,12 +482,14 @@ public:
 
             emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
         }
+        // sd::ggml_graph_cut::mark_graph_cut(emb, "unet.prelude", "emb");
 
         // input_blocks
         std::vector<ggml_tensor*> hs;
 
         // input block 0
         auto h = input_blocks_0_0->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(h, "unet.input_blocks.0", "h");
 
         ggml_set_name(h, "bench-start");
         hs.push_back(h);
@@ -505,6 +507,7 @@ public:
                     std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
                     h                = attention_layer_forward(name, ctx, h, context, num_video_frames);  // [N, mult*model_channels, h, w]
                 }
+                sd::ggml_graph_cut::mark_graph_cut(h, "unet.input_blocks." + std::to_string(input_block_idx), "h");
                 hs.push_back(h);
             }
             if (tiny_unet) {
@@ -518,6 +521,7 @@ public:
                 auto block       = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
 
                 h = block->forward(ctx, h);  // [N, mult*model_channels, h/(2^(i+1)), w/(2^(i+1))]
+                // sd::ggml_graph_cut::mark_graph_cut(h, "unet.input_blocks." + std::to_string(input_block_idx), "h");
                 hs.push_back(h);
             }
         }
@@ -531,6 +535,7 @@ public:
                 h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
             }
         }
+        sd::ggml_graph_cut::mark_graph_cut(h, "unet.middle_block", "h");
         if (controls.size() > 0) {
             auto cs = ggml_ext_scale(ctx->ggml_ctx, controls[controls.size() - 1], control_strength, true);
             h       = ggml_add(ctx->ggml_ctx, h, cs);  // middle control
@@ -581,6 +586,7 @@ public:
                 }
 
                 output_block_idx += 1;
+                sd::ggml_graph_cut::mark_graph_cut(h, "unet.output_blocks." + std::to_string(output_block_idx - 1), "h");
             }
         }
 

src/upscaler.cpp

@@ -1,50 +1,31 @@
-#include "esrgan.hpp"
+#include "upscaler.h"
 #include "ggml_extend.hpp"
 #include "model.h"
 #include "stable-diffusion.h"
 #include "util.h"
 
-struct UpscalerGGML {
-    ggml_backend_t backend    = nullptr;  // general backend
-    ggml_type model_data_type = GGML_TYPE_F16;
-    std::shared_ptr<ESRGAN> esrgan_upscaler;
-    std::string esrgan_path;
-    int n_threads;
-    bool direct   = false;
-    int tile_size = 128;
-
-    UpscalerGGML(int n_threads,
-                 bool direct   = false,
-                 int tile_size = 128)
+UpscalerGGML::UpscalerGGML(int n_threads,
+                           bool direct,
+                           int tile_size)
     : n_threads(n_threads),
       direct(direct),
       tile_size(tile_size) {
-    }
+}
 
-    bool load_from_file(const std::string& esrgan_path,
+void UpscalerGGML::set_max_graph_vram_bytes(size_t max_vram_bytes) {
+    max_graph_vram_bytes = max_vram_bytes;
+    if (esrgan_upscaler) {
+        esrgan_upscaler->set_max_graph_vram_bytes(max_vram_bytes);
+    }
+}
+
+bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
                                   bool offload_params_to_cpu,
                                   int n_threads) {
     ggml_log_set(ggml_log_callback_default, nullptr);
-#ifdef SD_USE_CUDA
-        LOG_DEBUG("Using CUDA backend");
-        backend = ggml_backend_cuda_init(0);
-#endif
-#ifdef SD_USE_METAL
-        LOG_DEBUG("Using Metal backend");
-        backend = ggml_backend_metal_init();
-#endif
-#ifdef SD_USE_VULKAN
-        LOG_DEBUG("Using Vulkan backend");
-        backend = ggml_backend_vk_init(0);
-#endif
-#ifdef SD_USE_OPENCL
-        LOG_DEBUG("Using OpenCL backend");
-        backend = ggml_backend_opencl_init();
-#endif
-#ifdef SD_USE_SYCL
-        LOG_DEBUG("Using SYCL backend");
-        backend = ggml_backend_sycl_init(0);
-#endif
+
+    backend = sd_get_default_backend();
+
     ModelLoader model_loader;
     if (!model_loader.init_from_file_and_convert_name(esrgan_path)) {
         LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
@@ -56,6 +37,7 @@ struct UpscalerGGML {
     }
     LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
     esrgan_upscaler = std::make_shared<ESRGAN>(backend, offload_params_to_cpu, tile_size, model_loader.get_tensor_storage_map());
+    esrgan_upscaler->set_max_graph_vram_bytes(max_graph_vram_bytes);
     if (direct) {
         esrgan_upscaler->set_conv2d_direct_enabled(true);
     }
@@ -63,9 +45,9 @@ struct UpscalerGGML {
         return false;
     }
     return true;
-    }
+}
 
-    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor) {
+sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_tensor) {
     sd::Tensor<float> upscaled;
     if (tile_size <= 0 || (input_tensor.shape()[0] <= tile_size && input_tensor.shape()[1] <= tile_size)) {
         upscaled = esrgan_upscaler->compute(n_threads, input_tensor);
@@ -96,9 +78,9 @@ struct UpscalerGGML {
         return {};
     }
     return upscaled;
-    }
+}
 
-    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
+sd_image_t UpscalerGGML::upscale(sd_image_t input_image, uint32_t upscale_factor) {
     // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
     sd_image_t upscaled_image = {0, 0, 0, nullptr};
     int output_width          = (int)input_image.width * esrgan_upscaler->scale;
@@ -118,8 +100,7 @@ struct UpscalerGGML {
     LOG_INFO("input_image_tensor upscaled, taking %.2fs", (t3 - t0) / 1000.0f);
     upscaled_image = upscaled_data;
     return upscaled_image;
-    }
-};
+}
 
 struct upscaler_ctx_t {
     UpscalerGGML* upscaler = nullptr;

src/upscaler.h

@@ -0,0 +1,33 @@
+#ifndef __SD_UPSCALER_H__
+#define __SD_UPSCALER_H__
+
+#include "esrgan.hpp"
+#include "stable-diffusion.h"
+#include "tensor.hpp"
+
+#include <memory>
+#include <string>
+
+struct UpscalerGGML {
+    ggml_backend_t backend    = nullptr;  // general backend
+    ggml_type model_data_type = GGML_TYPE_F16;
+    std::shared_ptr<ESRGAN> esrgan_upscaler;
+    std::string esrgan_path;
+    int n_threads;
+    bool direct                 = false;
+    int tile_size               = 128;
+    size_t max_graph_vram_bytes = 0;
+
+    UpscalerGGML(int n_threads,
+                 bool direct   = false,
+                 int tile_size = 128);
+
+    bool load_from_file(const std::string& esrgan_path,
+                        bool offload_params_to_cpu,
+                        int n_threads);
+    void set_max_graph_vram_bytes(size_t max_vram_bytes);
+    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor);
+    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor);
+};
+
+#endif  // __SD_UPSCALER_H__

src/util.cpp

@@ -23,8 +23,9 @@
 #include <unistd.h>
 #endif
 
-#include "ggml-cpu.h"
+#include "ggml-backend.h"
 #include "ggml.h"
+#include "ggml_extend_backend.hpp"
 #include "stable-diffusion.h"
 
 bool ends_with(const std::string& str, const std::string& ending) {
@@ -119,10 +120,10 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
         filename.c_str(),
         GENERIC_READ,
         FILE_SHARE_READ,
-        NULL,
+        nullptr,
         OPEN_EXISTING,
         FILE_ATTRIBUTE_NORMAL,
-        NULL);
+        nullptr);
 
     if (file_handle == INVALID_HANDLE_VALUE) {
         return nullptr;
@@ -136,16 +137,16 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
 
     file_size = static_cast<size_t>(size.QuadPart);
 
-    HANDLE mapping_handle = CreateFileMapping(file_handle, NULL, PAGE_READONLY, 0, 0, NULL);
+    HANDLE mapping_handle = CreateFileMapping(file_handle, nullptr, PAGE_READONLY, 0, 0, nullptr);
 
-    if (mapping_handle == NULL) {
+    if (mapping_handle == nullptr) {
         CloseHandle(file_handle);
         return nullptr;
     }
 
     mapped_data = MapViewOfFile(mapping_handle, FILE_MAP_READ, 0, 0, file_size);
 
-    if (mapped_data == NULL) {
+    if (mapped_data == nullptr) {
         CloseHandle(mapping_handle);
         CloseHandle(file_handle);
         return nullptr;
@@ -203,7 +204,7 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
 
     size_t file_size = sb.st_size;
 
-    void* mapped_data = mmap(NULL, file_size, PROT_READ, mmap_flags, file_descriptor, 0);
+    void* mapped_data = mmap(nullptr, file_size, PROT_READ, mmap_flags, file_descriptor, 0);
 
     close(file_descriptor);
 
@@ -495,26 +496,6 @@ sd_progress_cb_t sd_get_progress_callback() {
 void* sd_get_progress_callback_data() {
     return sd_progress_cb_data;
 }
-const char* sd_get_system_info() {
-    static char buffer[1024];
-    std::stringstream ss;
-    ss << "System Info: \n";
-    ss << "    SSE3 = " << ggml_cpu_has_sse3() << " | ";
-    ss << "    AVX = " << ggml_cpu_has_avx() << " | ";
-    ss << "    AVX2 = " << ggml_cpu_has_avx2() << " | ";
-    ss << "    AVX512 = " << ggml_cpu_has_avx512() << " | ";
-    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << " | ";
-    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << " | ";
-    ss << "    FMA = " << ggml_cpu_has_fma() << " | ";
-    ss << "    NEON = " << ggml_cpu_has_neon() << " | ";
-    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << " | ";
-    ss << "    F16C = " << ggml_cpu_has_f16c() << " | ";
-    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << " | ";
-    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << " | ";
-    ss << "    VSX = " << ggml_cpu_has_vsx() << " | ";
-    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
-    return buffer;
-}
 
 sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index) {
     const auto& shape = tensor.shape();
@@ -524,17 +505,7 @@ sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index)
     int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
     uint8_t* data = (uint8_t*)malloc(static_cast<size_t>(width * height * channel));
     GGML_ASSERT(data != nullptr);
-
-    for (int iw = 0; iw < width; ++iw) {
-        for (int ih = 0; ih < height; ++ih) {
-            for (int ic = 0; ic < channel; ++ic) {
-                float value                            = shape.size() == 5 ? tensor.index(iw, ih, frame_index, ic, 0)
-                                                                           : tensor.index(iw, ih, ic, frame_index);
-                value                                  = std::clamp(value, 0.0f, 1.0f);
-                data[(ih * width + iw) * channel + ic] = static_cast<uint8_t>(std::round(value * 255.0f));
-            }
-        }
-    }
+    preprocessing_tensor_frame_to_sd_image(tensor, frame_index, data);
     return {
         static_cast<uint32_t>(width),
         static_cast<uint32_t>(height),
@@ -718,3 +689,100 @@ std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::str
 
     return res;
 }
+
+// test if the backend is a specific one, e.g. "CUDA", "ROCm", "Vulkan" etc.
+bool sd_backend_is(ggml_backend_t backend, const std::string& name) {
+    if (!backend) {
+        return false;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (!dev)
+        return false;
+    std::string dev_name = ggml_backend_dev_name(dev);
+    return dev_name.find(name) != std::string::npos;
+}
+
+ggml_backend_t sd_get_default_backend() {
+    ggml_backend_load_all_once();
+    static std::once_flag once;
+    std::call_once(once, []() {
+        size_t dev_count = ggml_backend_dev_count();
+        if (dev_count == 0) {
+            LOG_ERROR("No devices found!");
+        } else {
+            LOG_DEBUG("Found %zu backend devices:", dev_count);
+            for (size_t i = 0; i < dev_count; ++i) {
+                auto dev = ggml_backend_dev_get(i);
+                LOG_DEBUG("#%zu: %s", i, ggml_backend_dev_name(dev));
+            }
+        }
+    });
+    ggml_backend_t backend   = nullptr;
+    const char* SD_VK_DEVICE = getenv("SD_VK_DEVICE");
+    if (SD_VK_DEVICE != nullptr) {
+        std::string sd_vk_device_str = SD_VK_DEVICE;
+        try {
+            unsigned long long device  = std::stoull(sd_vk_device_str);
+            std::string vk_device_name = "Vulkan" + std::to_string(device);
+            if (backend_name_exists(vk_device_name)) {
+                LOG_INFO("Selecting %s as main device by env var SD_VK_DEVICE", vk_device_name.c_str());
+                backend = init_named_backend(vk_device_name);
+                if (!backend) {
+                    LOG_WARN("Device %s requested by SD_VK_DEVICE failed to init. Falling back to the default device.", vk_device_name.c_str());
+                }
+            } else {
+                LOG_WARN("Device %s requested by SD_VK_DEVICE was not found. Falling back to the default device.", vk_device_name.c_str());
+            }
+        } catch (const std::invalid_argument&) {
+            LOG_WARN("SD_VK_DEVICE environment variable is not a valid integer (%s). Falling back to the default device.", SD_VK_DEVICE);
+        } catch (const std::out_of_range&) {
+            LOG_WARN("SD_VK_DEVICE environment variable value is out of range for `unsigned long long` type (%s). Falling back to the default device.", SD_VK_DEVICE);
+        }
+    }
+
+    if (!backend) {
+        std::string dev_name = get_default_backend_name();
+        backend              = init_named_backend(dev_name);
+        if (!backend && !dev_name.empty()) {
+            LOG_WARN("device %s failed to init", dev_name.c_str());
+        }
+    }
+
+    if (!backend) {
+        LOG_WARN("loading CPU backend");
+        backend = ggml_backend_cpu_init();
+    }
+
+    if (ggml_backend_is_cpu(backend)) {
+        LOG_DEBUG("Using CPU backend");
+    }
+
+    return backend;
+}
+
+// namespace is needed to avoid conflicts with ggml_backend_extend.hpp
+namespace ggml_cpu {
+#include "ggml-cpu.h"
+}
+
+const char* sd_get_system_info() {
+    using namespace ggml_cpu;
+    static char buffer[1024];
+    std::stringstream ss;
+    ss << "System Info: \n";
+    ss << "    SSE3 = " << ggml_cpu_has_sse3() << " | ";
+    ss << "    AVX = " << ggml_cpu_has_avx() << " | ";
+    ss << "    AVX2 = " << ggml_cpu_has_avx2() << " | ";
+    ss << "    AVX512 = " << ggml_cpu_has_avx512() << " | ";
+    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << " | ";
+    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << " | ";
+    ss << "    FMA = " << ggml_cpu_has_fma() << " | ";
+    ss << "    NEON = " << ggml_cpu_has_neon() << " | ";
+    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << " | ";
+    ss << "    F16C = " << ggml_cpu_has_f16c() << " | ";
+    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << " | ";
+    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << " | ";
+    ss << "    VSX = " << ggml_cpu_has_vsx() << " | ";
+    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
+    return buffer;
+}

src/util.h

@@ -6,6 +6,7 @@
 #include <string>
 #include <vector>
 
+#include "ggml-backend.h"
 #include "stable-diffusion.h"
 #include "tensor.hpp"
 
@@ -82,6 +83,10 @@ int sd_get_preview_interval();
 bool sd_should_preview_denoised();
 bool sd_should_preview_noisy();
 
+// test if the backend is a specific one, e.g. "CUDA", "ROCm", "Vulkan" etc.
+bool sd_backend_is(ggml_backend_t backend, const std::string& name);
+ggml_backend_t sd_get_default_backend();
+
 #define LOG_DEBUG(format, ...) log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #define LOG_INFO(format, ...) log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #define LOG_WARN(format, ...) log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)

src/vae.hpp

@@ -142,9 +142,10 @@ public:
                                    "vae encode compute failed while processing a tile");
         } else {
             output = _compute(n_threads, input, false);
-            free_compute_buffer();
         }
 
+        free_compute_buffer();
+
         if (output.empty()) {
             LOG_ERROR("vae encode compute failed");
             return {};

src/wan.hpp

@@ -692,6 +692,7 @@ namespace WAN {
             } else {
                 x = conv1->forward(ctx, x);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.prelude", "x");
 
             // downsamples
             std::vector<int64_t> dims = {dim};
@@ -717,12 +718,14 @@ namespace WAN {
                         x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
                     }
                 }
+                // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.down." + std::to_string(i), "x");
             }
 
             // middle
             x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
             x = middle_1->forward(ctx, x, b);
             x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.mid", "x");
 
             // head
             x = head_0->forward(ctx, x);
@@ -863,11 +866,13 @@ namespace WAN {
             } else {
                 x = conv1->forward(ctx, x);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.prelude", "x");
 
             // middle
             x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
             x = middle_1->forward(ctx, x, b);
             x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.mid", "x");
 
             // upsamples
             std::vector<int64_t> dims = {dim_mult[dim_mult.size() - 1] * dim};
@@ -893,6 +898,7 @@ namespace WAN {
                         x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
                     }
                 }
+                // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.up." + std::to_string(i), "x");
             }
 
             // head
@@ -1031,6 +1037,7 @@ namespace WAN {
             if (wan2_2) {
                 x = patchify(ctx->ggml_ctx, x, 2, b);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encode.prelude", "x");
 
             auto encoder = std::dynamic_pointer_cast<Encoder3d>(blocks["encoder"]);
             auto conv1   = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
@@ -1051,6 +1058,7 @@ namespace WAN {
             }
             out     = conv1->forward(ctx, out);
             auto mu = ggml_ext_chunk(ctx->ggml_ctx, out, 2, 3)[0];
+            // sd::ggml_graph_cut::mark_graph_cut(mu, "wan_vae.encode.final", "mu");
             clear_cache();
             return mu;
         }
@@ -1068,6 +1076,7 @@ namespace WAN {
 
             int64_t iter_ = z->ne[2];
             auto x        = conv2->forward(ctx, z);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decode.prelude", "x");
             ggml_tensor* out;
             for (int i = 0; i < iter_; i++) {
                 _conv_idx = 0;
@@ -1083,6 +1092,7 @@ namespace WAN {
             if (wan2_2) {
                 out = unpatchify(ctx->ggml_ctx, out, 2, b);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode.final", "out");
             clear_cache();
             return out;
         }
@@ -1098,12 +1108,14 @@ namespace WAN {
             auto conv2   = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv2"]);
 
             auto x = conv2->forward(ctx, z);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decode_partial.prelude", "x");
             auto in   = ggml_ext_slice(ctx->ggml_ctx, x, 2, i, i + 1);  // [b*c, 1, h, w]
             _conv_idx = 0;
             auto out  = decoder->forward(ctx, in, b, _feat_map, _conv_idx, i);
             if (wan2_2) {
                 out = unpatchify(ctx->ggml_ctx, out, 2, b);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode_partial.final", "out");
             return out;
         }
     };
@@ -1984,6 +1996,13 @@ namespace WAN {
                 c = ggml_reshape_3d(ctx->ggml_ctx, c, c->ne[0] * c->ne[1] * c->ne[2], c->ne[3] / N, N);  // [N, dim, t_len*h_len*w_len]
                 c = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, c, 1, 0, 2, 3));  // [N, t_len*h_len*w_len, dim]
             }
+            sd::ggml_graph_cut::mark_graph_cut(x, "wan.prelude", "x");
+            // sd::ggml_graph_cut::mark_graph_cut(e, "wan.prelude", "e");
+            // sd::ggml_graph_cut::mark_graph_cut(e0, "wan.prelude", "e0");
+            // sd::ggml_graph_cut::mark_graph_cut(context, "wan.prelude", "context");
+            if (c != nullptr) {
+                sd::ggml_graph_cut::mark_graph_cut(c, "wan.prelude", "c");
+            }
 
             auto x_orig = x;
 
@@ -2004,6 +2023,10 @@ namespace WAN {
                     c_skip      = ggml_ext_scale(ctx->ggml_ctx, c_skip, vace_strength);
                     x           = ggml_add(ctx->ggml_ctx, x, c_skip);
                 }
+                sd::ggml_graph_cut::mark_graph_cut(x, "wan.blocks." + std::to_string(i), "x");
+                if (c != nullptr) {
+                    sd::ggml_graph_cut::mark_graph_cut(c, "wan.blocks." + std::to_string(i), "c");
+                }
             }
 
             x = head->forward(ctx, x, e);  // [N, t_len*h_len*w_len, pt*ph*pw*out_dim]

src/z_image.hpp

@@ -31,10 +31,6 @@ namespace ZImage {
             : head_dim(head_dim), num_heads(num_heads), num_kv_heads(num_kv_heads), qk_norm(qk_norm) {
             blocks["qkv"] = std::make_shared<Linear>(hidden_size, (num_heads + num_kv_heads * 2) * head_dim, false);
             float scale   = 1.f;
-#if GGML_USE_HIP
-            // Prevent NaN issues with certain ROCm setups
-            scale = 1.f / 16.f;
-#endif
             blocks["out"] = std::make_shared<Linear>(num_heads * head_dim, hidden_size, false, false, false, scale);
             if (qk_norm) {
                 blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim);
@@ -52,6 +48,10 @@ namespace ZImage {
             auto qkv_proj   = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
             auto out_proj   = std::dynamic_pointer_cast<Linear>(blocks["out"]);
 
+            if (sd_backend_is(ctx->backend, "ROCm")) {
+                out_proj->set_scale(1.f / 16.f);
+            }
+
             auto qkv = qkv_proj->forward(ctx, x);                                                                            // [N, n_token, (num_heads + num_kv_heads*2)*head_dim]
             qkv      = ggml_reshape_4d(ctx->ggml_ctx, qkv, head_dim, num_heads + num_kv_heads * 2, qkv->ne[1], qkv->ne[2]);  // [N, n_token, num_heads + num_kv_heads*2, head_dim]
 
@@ -115,9 +115,7 @@ namespace ZImage {
 
             bool force_prec_f32 = false;
             float scale         = 1.f / 128.f;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
+
             // The purpose of the scale here is to prevent NaN issues in certain situations.
             // For example, when using CUDA but the weights are k-quants.
             blocks["w2"] = std::make_shared<Linear>(hidden_dim, dim, false, false, force_prec_f32, scale);
@@ -129,6 +127,10 @@ namespace ZImage {
             auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
             auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
 
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                w2->set_force_prec_f32(true);
+            }
+
             auto x1 = w1->forward(ctx, x);
             auto x3 = w3->forward(ctx, x);
             x       = ggml_swiglu_split(ctx->ggml_ctx, x1, x3);
@@ -369,6 +371,9 @@ namespace ZImage {
 
             auto txt = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context));  // [N, n_txt_token, hidden_size]
             auto img = x_embedder->forward(ctx, x);                                          // [N, n_img_token, hidden_size]
+            sd::ggml_graph_cut::mark_graph_cut(txt, "z_image.prelude", "txt");
+            sd::ggml_graph_cut::mark_graph_cut(img, "z_image.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(t_emb, "z_image.prelude", "t_emb");
 
             int64_t n_txt_pad_token = Rope::bound_mod(static_cast<int>(n_txt_token), SEQ_MULTI_OF);
             if (n_txt_pad_token > 0) {
@@ -391,20 +396,24 @@ namespace ZImage {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
 
                 txt = block->forward(ctx, txt, txt_pe, nullptr, nullptr);
+                sd::ggml_graph_cut::mark_graph_cut(txt, "z_image.context_refiner." + std::to_string(i), "txt");
             }
 
             for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
 
                 img = block->forward(ctx, img, img_pe, nullptr, t_emb);
+                sd::ggml_graph_cut::mark_graph_cut(img, "z_image.noise_refiner." + std::to_string(i), "img");
             }
 
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_txt_pad_token + n_img_token + n_img_pad_token, hidden_size]
+            sd::ggml_graph_cut::mark_graph_cut(txt_img, "z_image.prelude", "txt_img");
 
             for (int i = 0; i < z_image_params.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["layers." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, pe, nullptr, t_emb);
+                sd::ggml_graph_cut::mark_graph_cut(txt_img, "z_image.layers." + std::to_string(i), "txt_img");
             }
 
             txt_img = final_layer->forward(ctx, txt_img, t_emb);  // [N, n_txt_token + n_txt_pad_token + n_img_token + n_img_pad_token, ph*pw*C]