Is it expected that the --cache-preset option doesn't work? #1188

masamaru-san · 2026-01-12T13:14:42Z

masamaru-san
Jan 12, 2026

❓Question:

While experimenting with the --cache-mode cache-dit option on sd-cli, I noticed that the effects of --cache-preset were not being applied. Is this the expected behavior?

Test condition, results and logs

My environment:

Item	Value
Processor:	AMD Ryzen 7 5700U with Radeon Graphics (1.80 GHz)
Physical RAM:	64.0 GB (63.3 GB available)
OS:	Windows 11 Home 25H2 [Version 10.0.26200.7462]
Used binary:	sd-master-885e62e-bin-win-vulkan-x64.zip

Test script (powershell 7):

& {
  cmd.exe /c ver
  vulkaninfo.exe --summary

  "`nPreset MODE: none"
  .\sd-cli.exe `
    -m 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors' `
    --clip_l 'E:\AI\models\sd3_0_3_5\clip_l.safetensors' `
    --clip_g 'E:\AI\models\sd3_0_3_5\clip_g.safetensors' `
    --vae 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors' `
    --t5xxl 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors' `
    -p 'a cat' --cfg-scale 4.5 --sampling-method euler `
    --diffusion-conv-direct --vae-conv-direct `
    -t 15 --mmap --cache-mode 'cache-dit' `
    -v -o '.\cat_preset_none.png'

  "`nPreset MODE: ultra"
  .\sd-cli.exe `
    -m 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors' `
    --clip_l 'E:\AI\models\sd3_0_3_5\clip_l.safetensors' `
    --clip_g 'E:\AI\models\sd3_0_3_5\clip_g.safetensors' `
    --vae 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors' `
    --t5xxl 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors' `
    -p 'a cat' --cfg-scale 4.5 --sampling-method euler `
    --diffusion-conv-direct --vae-conv-direct `
    -t 15 --mmap --cache-mode 'cache-dit' --cache-preset ultra `
    -v -o '.\cat_preset_ultra.png'

  "`nPreset MODE: fast"
  .\sd-cli.exe `
    -m 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors' `
    --clip_l 'E:\AI\models\sd3_0_3_5\clip_l.safetensors' `
    --clip_g 'E:\AI\models\sd3_0_3_5\clip_g.safetensors' `
    --vae 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors' `
    --t5xxl 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors' `
    -p 'a cat' --cfg-scale 4.5 --sampling-method euler `
    --diffusion-conv-direct --vae-conv-direct `
    -t 15 --mmap --cache-mode 'cache-dit' --cache-preset fast `
    -v -o '.\cat_preset_fast.png'

  "`nPreset MODE: medium"
  .\sd-cli.exe `
    -m 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors' `
    --clip_l 'E:\AI\models\sd3_0_3_5\clip_l.safetensors' `
    --clip_g 'E:\AI\models\sd3_0_3_5\clip_g.safetensors' `
    --vae 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors' `
    --t5xxl 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors' `
    -p 'a cat' --cfg-scale 4.5 --sampling-method euler `
    --diffusion-conv-direct --vae-conv-direct `
    -t 15 --mmap --cache-mode 'cache-dit' --cache-preset medium `
    -v -o '.\cat_preset_medium.png'

  "`nPreset MODE: slow"
  .\sd-cli.exe `
    -m 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors' `
    --clip_l 'E:\AI\models\sd3_0_3_5\clip_l.safetensors' `
    --clip_g 'E:\AI\models\sd3_0_3_5\clip_g.safetensors' `
    --vae 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors' `
    --t5xxl 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors' `
    -p 'a cat' --cfg-scale 4.5 --sampling-method euler `
    --diffusion-conv-direct --vae-conv-direct `
    -t 15 --mmap --cache-mode 'cache-dit' --cache-preset slow `
    -v -o '.\cat_preset_slow.png'
}

Results and Logs

Summary:

It seems the same default parameters are applied under all conditions. All generated images are exactly the same.

   ...

  cache_mode: "cache-dit",
  cache_option: "",
  cache: enabled (threshold=1, start=0.15, end=0.95),

   ...

[INFO ] stable-diffusion.cpp:1742 - CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
[DEBUG] ggml_extend.hpp:1733 - mmdit compute buffer size: 181.64 MB(VRAM)
  |==================================================| 20/20 - 6.25s/it
[INFO ] stable-diffusion.cpp:2219 - CacheDIT skipped 4/20 steps (1.25x estimated speedup), accum_diff: 0.2213

default	ultra	fast	medium	slow

All logs:

cmd.exe /c ver

Microsoft Windows [Version 10.0.26200.7462]

vulkaninfo

==========
VULKANINFO
==========

Vulkan Instance Version: 1.4.335


Instance Extensions: count = 13
-------------------------------
VK_EXT_debug_report                    : extension revision 10
VK_EXT_debug_utils                     : extension revision 2
VK_EXT_swapchain_colorspace            : extension revision 4
VK_KHR_device_group_creation           : extension revision 1
VK_KHR_external_fence_capabilities     : extension revision 1
VK_KHR_external_memory_capabilities    : extension revision 1
VK_KHR_external_semaphore_capabilities : extension revision 1
VK_KHR_get_physical_device_properties2 : extension revision 2
VK_KHR_get_surface_capabilities2       : extension revision 1
VK_KHR_portability_enumeration         : extension revision 1
VK_KHR_surface                         : extension revision 25
VK_KHR_win32_surface                   : extension revision 6
VK_LUNARG_direct_driver_loading        : extension revision 1

Instance Layers: count = 10
---------------------------
VK_LAYER_AMD_switchable_graphics  AMD switchable graphics layer                                                                                     1.3.260  version 1
VK_LAYER_KHRONOS_profiles         Khronos Profiles layer                                                                                            1.4.335  version 1
VK_LAYER_KHRONOS_shader_object    Khronos Shader object layer                                                                                       1.4.335  version 1
VK_LAYER_KHRONOS_synchronization2 Khronos Synchronization2 layer                                                                                    1.4.335  version 1
VK_LAYER_KHRONOS_validation       Khronos Validation Layer                                                                                          1.4.335  version 1
VK_LAYER_LUNARG_api_dump          LunarG API dump layer                                                                                             1.4.335  version 2
VK_LAYER_LUNARG_crash_diagnostic  Crash Diagnostic Layer is a crash/hang debugging tool that helps determines GPU progress in a Vulkan application. 1.4.335  version 1
VK_LAYER_LUNARG_gfxreconstruct    GFXReconstruct Capture Layer Version 1.0.5                                                                        1.4.335  version 4194309
VK_LAYER_LUNARG_monitor           Execution Monitoring Layer                                                                                        1.4.335  version 1
VK_LAYER_LUNARG_screenshot        LunarG image capture layer                                                                                        1.4.335  version 1

Devices:
========
GPU0:
        apiVersion         = 1.3.260
        driverVersion      = 2.0.279
        vendorID           = 0x1002
        deviceID           = 0x164c
        deviceType         = PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
        deviceName         = AMD Radeon(TM) Graphics
        driverID           = DRIVER_ID_AMD_PROPRIETARY
        driverName         = AMD proprietary driver
        driverInfo         = 25.8.1 (AMD proprietary shader compiler)
        conformanceVersion = 1.3.3.1
        deviceUUID         = 00000000-0400-0000-0000-000000000000
        driverUUID         = 414d442d-5749-4e2d-4452-560000000000

Preset MODE: none (default)

Preset MODE: none
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 885e62e
[DEBUG] main.cpp:501  - System Info:
    SSE3 = 1 |     AVX = 1 |     AVX2 = 1 |     AVX512 = 0 |     AVX512_VBMI = 0 |     AVX512_VNNI = 0 |     FMA = 1 |     NEON = 0 |     ARM_FMA = 0 |     F16C = 1 |     FP16_VA = 0 |     WASM_SIMD = 0 |     VSX = 0 |
[DEBUG] main.cpp:502  - SDCliParams {
  mode: img_gen,
  output_path: ".\cat_preset_none.png",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false
}
[DEBUG] main.cpp:503  - SDContextParams {
  n_threads: 15,
  model_path: "E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors",
  clip_l_path: "E:\AI\models\sd3_0_3_5\clip_l.safetensors",
  clip_g_path: "E:\AI\models\sd3_0_3_5\clip_g.safetensors",
  clip_vision_path: "",
  t5xxl_path: "E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors",
  llm_path: "",
  llm_vision_path: "",
  diffusion_model_path: "",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "",
  photo_maker_path: "",
  rng_type: cuda,
  sampler_rng_type: NONE,
  flow_shift: INF
  offload_params_to_cpu: false,
  enable_mmap: true,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  diffusion_flash_attn: false,
  diffusion_conv_direct: true,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: auto,
  vae_tiling_params: { 0, 0, 0, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "a cat",
  negative_prompt: "",
  clip_skip: -1,
  width: 512,
  height: 512,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 4.50, img_cfg: 4.50, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: euler, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  custom_sigmas: [],
  cache_mode: "cache-dit",
  cache_option: "",
  cache: enabled (threshold=1, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 42,
  upscale_repeats: 1,
  upscale_tile_size: 128,
}
[DEBUG] stable-diffusion.cpp:169  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:190  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:248  - loading model from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors', prefix = ''
[INFO ] stable-diffusion.cpp:271  - loading clip_l from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_l.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors', prefix = 'text_encoders.clip_l.transformer.'
[INFO ] stable-diffusion.cpp:279  - loading clip_g from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_g.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors', prefix = 'text_encoders.clip_g.transformer.'
[INFO ] stable-diffusion.cpp:295  - loading t5xxl from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors', prefix = 'text_encoders.t5xxl.transformer.'
[INFO ] stable-diffusion.cpp:316  - loading vae from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:332  - Version: SD3.x
[INFO ] stable-diffusion.cpp:360  - Weight type stat:                      f16: 1597 |    bf16: 244
[INFO ] stable-diffusion.cpp:361  - Conditioner weight type stat:          f16: 932
[INFO ] stable-diffusion.cpp:362  - Diffusion model weight type stat:      f16: 665
[INFO ] stable-diffusion.cpp:363  - VAE weight type stat:                 bf16: 244
[DEBUG] stable-diffusion.cpp:365  - ggml tensor size = 400 bytes
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[INFO ] mmdit.hpp:690  - MMDiT layers: 24 (including 13 MMDiT-x layers)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  235.06 MB(VRAM) (196 tensors)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  1329.29 MB(VRAM) (517 tensors)
[DEBUG] ggml_extend.hpp:1921 - t5 params backend buffer size =  9083.77 MB(VRAM) (219 tensors)
[DEBUG] ggml_extend.hpp:1921 - mmdit params backend buffer size =  5172.83 MB(VRAM) (665 tensors)
[INFO ] stable-diffusion.cpp:616  - Using Conv2d direct in the vae model
[DEBUG] ggml_extend.hpp:1921 - vae params backend buffer size =  94.57 MB(VRAM) (138 tensors)
[DEBUG] stable-diffusion.cpp:741  - loading weights
[DEBUG] model.cpp:1351 - using 15 threads for model loading
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================>                               | 665/1841 - 412.28it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_l.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=======================>                          | 861/1841 - 333.20it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_g.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=====================================>            | 1378/1841 - 449.59it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |===========================================>      | 1597/1841 - 319.34it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================================================| 1841/1841 - 279.07it/s
[INFO ] model.cpp:1599 - loading tensors completed, taking 6.62s (process: 0.00s, read: 1.92s, memcpy: 0.00s, convert: 0.04s, copy_to_backend: 0.63s)
[DEBUG] stable-diffusion.cpp:776  - finished loaded file
[INFO ] stable-diffusion.cpp:848  - total params memory size = 15915.53MB (VRAM 15915.53MB, RAM 0.00MB): text_encoders 10648.12MB(VRAM), diffusion_model 5172.83MB(VRAM), vae 94.57MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
[INFO ] stable-diffusion.cpp:917  - running in FLOW mode
[DEBUG] stable-diffusion.cpp:3460 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3494 - sampling using Euler method
[INFO ] denoiser.hpp:403  - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3621 - TXT2IMG
[DEBUG] conditioner.hpp:810  - parse 'a cat' to [['a cat', 1], ]
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1732 ms
[DEBUG] conditioner.hpp:810  - parse '' to [['', 1], ]
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1713 ms
[INFO ] stable-diffusion.cpp:3238 - get_learned_condition completed, taking 3451 ms
[INFO ] stable-diffusion.cpp:3349 - generating image: 1/1 - seed 42
[INFO ] stable-diffusion.cpp:1742 - CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
[DEBUG] ggml_extend.hpp:1733 - mmdit compute buffer size: 181.64 MB(VRAM)
  |==================================================| 20/20 - 6.25s/it
[INFO ] stable-diffusion.cpp:2219 - CacheDIT skipped 4/20 steps (1.25x estimated speedup), accum_diff: 0.2213
[INFO ] stable-diffusion.cpp:3391 - sampling completed, taking 125.10s
[INFO ] stable-diffusion.cpp:3402 - generating 1 latent images completed, taking 126.81s
[INFO ] stable-diffusion.cpp:3405 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1733 - vae compute buffer size: 704.25 MB(VRAM)
[DEBUG] stable-diffusion.cpp:2672 - computing vae decode graph completed, taking 8.14s
[INFO ] stable-diffusion.cpp:3415 - latent 1 decoded, taking 8.14s
[INFO ] stable-diffusion.cpp:3419 - decode_first_stage completed, taking 8.14s
[INFO ] stable-diffusion.cpp:3729 - generate_image completed in 138.41s
[INFO ] main.cpp:421  - save result image 0 to '.\cat_preset_none.png' (success)

Preset MODE: ultra

Preset MODE: ultra
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 885e62e
[DEBUG] main.cpp:501  - System Info:
    SSE3 = 1 |     AVX = 1 |     AVX2 = 1 |     AVX512 = 0 |     AVX512_VBMI = 0 |     AVX512_VNNI = 0 |     FMA = 1 |     NEON = 0 |     ARM_FMA = 0 |     F16C = 1 |     FP16_VA = 0 |     WASM_SIMD = 0 |     VSX = 0 |
[DEBUG] main.cpp:502  - SDCliParams {
  mode: img_gen,
  output_path: ".\cat_preset_ultra.png",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false
}
[DEBUG] main.cpp:503  - SDContextParams {
  n_threads: 15,
  model_path: "E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors",
  clip_l_path: "E:\AI\models\sd3_0_3_5\clip_l.safetensors",
  clip_g_path: "E:\AI\models\sd3_0_3_5\clip_g.safetensors",
  clip_vision_path: "",
  t5xxl_path: "E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors",
  llm_path: "",
  llm_vision_path: "",
  diffusion_model_path: "",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "",
  photo_maker_path: "",
  rng_type: cuda,
  sampler_rng_type: NONE,
  flow_shift: INF
  offload_params_to_cpu: false,
  enable_mmap: true,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  diffusion_flash_attn: false,
  diffusion_conv_direct: true,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: auto,
  vae_tiling_params: { 0, 0, 0, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "a cat",
  negative_prompt: "",
  clip_skip: -1,
  width: 512,
  height: 512,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 4.50, img_cfg: 4.50, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: euler, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  custom_sigmas: [],
  cache_mode: "cache-dit",
  cache_option: "",
  cache: enabled (threshold=1, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 42,
  upscale_repeats: 1,
  upscale_tile_size: 128,
}
[DEBUG] stable-diffusion.cpp:169  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:190  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:248  - loading model from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors', prefix = ''
[INFO ] stable-diffusion.cpp:271  - loading clip_l from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_l.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors', prefix = 'text_encoders.clip_l.transformer.'
[INFO ] stable-diffusion.cpp:279  - loading clip_g from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_g.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors', prefix = 'text_encoders.clip_g.transformer.'
[INFO ] stable-diffusion.cpp:295  - loading t5xxl from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors', prefix = 'text_encoders.t5xxl.transformer.'
[INFO ] stable-diffusion.cpp:316  - loading vae from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:332  - Version: SD3.x
[INFO ] stable-diffusion.cpp:360  - Weight type stat:                      f16: 1597 |    bf16: 244
[INFO ] stable-diffusion.cpp:361  - Conditioner weight type stat:          f16: 932
[INFO ] stable-diffusion.cpp:362  - Diffusion model weight type stat:      f16: 665
[INFO ] stable-diffusion.cpp:363  - VAE weight type stat:                 bf16: 244
[DEBUG] stable-diffusion.cpp:365  - ggml tensor size = 400 bytes
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[INFO ] mmdit.hpp:690  - MMDiT layers: 24 (including 13 MMDiT-x layers)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  235.06 MB(VRAM) (196 tensors)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  1329.29 MB(VRAM) (517 tensors)
[DEBUG] ggml_extend.hpp:1921 - t5 params backend buffer size =  9083.77 MB(VRAM) (219 tensors)
[DEBUG] ggml_extend.hpp:1921 - mmdit params backend buffer size =  5172.83 MB(VRAM) (665 tensors)
[INFO ] stable-diffusion.cpp:616  - Using Conv2d direct in the vae model
[DEBUG] ggml_extend.hpp:1921 - vae params backend buffer size =  94.57 MB(VRAM) (138 tensors)
[DEBUG] stable-diffusion.cpp:741  - loading weights
[DEBUG] model.cpp:1351 - using 15 threads for model loading
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================>                               | 665/1841 - 432.94it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_l.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=======================>                          | 861/1841 - 348.72it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_g.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=====================================>            | 1378/1841 - 469.51it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |===========================================>      | 1597/1841 - 327.93it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================================================| 1841/1841 - 285.96it/s
[INFO ] model.cpp:1599 - loading tensors completed, taking 6.46s (process: 0.00s, read: 1.89s, memcpy: 0.00s, convert: 0.04s, copy_to_backend: 0.64s)
[DEBUG] stable-diffusion.cpp:776  - finished loaded file
[INFO ] stable-diffusion.cpp:848  - total params memory size = 15915.53MB (VRAM 15915.53MB, RAM 0.00MB): text_encoders 10648.12MB(VRAM), diffusion_model 5172.83MB(VRAM), vae 94.57MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
[INFO ] stable-diffusion.cpp:917  - running in FLOW mode
[DEBUG] stable-diffusion.cpp:3460 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3494 - sampling using Euler method
[INFO ] denoiser.hpp:403  - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3621 - TXT2IMG
[DEBUG] conditioner.hpp:810  - parse 'a cat' to [['a cat', 1], ]
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1732 ms
[DEBUG] conditioner.hpp:810  - parse '' to [['', 1], ]
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1703 ms
[INFO ] stable-diffusion.cpp:3238 - get_learned_condition completed, taking 3442 ms
[INFO ] stable-diffusion.cpp:3349 - generating image: 1/1 - seed 42
[INFO ] stable-diffusion.cpp:1742 - CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
[DEBUG] ggml_extend.hpp:1733 - mmdit compute buffer size: 181.64 MB(VRAM)
  |==================================================| 20/20 - 6.23s/it
[INFO ] stable-diffusion.cpp:2219 - CacheDIT skipped 4/20 steps (1.25x estimated speedup), accum_diff: 0.2213
[INFO ] stable-diffusion.cpp:3391 - sampling completed, taking 124.69s
[INFO ] stable-diffusion.cpp:3402 - generating 1 latent images completed, taking 126.35s
[INFO ] stable-diffusion.cpp:3405 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1733 - vae compute buffer size: 704.25 MB(VRAM)
[DEBUG] stable-diffusion.cpp:2672 - computing vae decode graph completed, taking 8.23s
[INFO ] stable-diffusion.cpp:3415 - latent 1 decoded, taking 8.23s
[INFO ] stable-diffusion.cpp:3419 - decode_first_stage completed, taking 8.24s
[INFO ] stable-diffusion.cpp:3729 - generate_image completed in 138.04s
[INFO ] main.cpp:421  - save result image 0 to '.\cat_preset_ultra.png' (success)

Preset MODE: fast

<img width="512" height="512" alt="cat_preset_fast" src="https://github.com/user-attachments/assets/bd78a11a-6efc-4d43-90e4-44423d92f864" />
<img width="512" height="512" alt="cat_preset_medium" src="https://github.com/user-attachments/assets/6a0f2973-328f-47f6-8aec-74c974641585" />
<img width="512" height="512" alt="cat_preset_none" src="https://github.com/user-attachments/assets/5f16e262-b85f-4e76-8eb7-75530b0c009f" />
<img width="512" height="512" alt="cat_preset_slow" src="https://github.com/user-attachments/assets/023d60d8-c40f-43e3-9651-b22edf8051ee" />
<img width="512" height="512" alt="cat_preset_ultra" src="https://github.com/user-attachments/assets/b3b98d14-464a-419b-8e37-75a268dba5c7" />

Preset MODE: fast
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 885e62e
[DEBUG] main.cpp:501  - System Info:
    SSE3 = 1 |     AVX = 1 |     AVX2 = 1 |     AVX512 = 0 |     AVX512_VBMI = 0 |     AVX512_VNNI = 0 |     FMA = 1 |     NEON = 0 |     ARM_FMA = 0 |     F16C = 1 |     FP16_VA = 0 |     WASM_SIMD = 0 |     VSX = 0 |
[DEBUG] main.cpp:502  - SDCliParams {
  mode: img_gen,
  output_path: ".\cat_preset_fast.png",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false
}
[DEBUG] main.cpp:503  - SDContextParams {
  n_threads: 15,
  model_path: "E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors",
  clip_l_path: "E:\AI\models\sd3_0_3_5\clip_l.safetensors",
  clip_g_path: "E:\AI\models\sd3_0_3_5\clip_g.safetensors",
  clip_vision_path: "",
  t5xxl_path: "E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors",
  llm_path: "",
  llm_vision_path: "",
  diffusion_model_path: "",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "",
  photo_maker_path: "",
  rng_type: cuda,
  sampler_rng_type: NONE,
  flow_shift: INF
  offload_params_to_cpu: false,
  enable_mmap: true,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  diffusion_flash_attn: false,
  diffusion_conv_direct: true,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: auto,
  vae_tiling_params: { 0, 0, 0, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "a cat",
  negative_prompt: "",
  clip_skip: -1,
  width: 512,
  height: 512,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 4.50, img_cfg: 4.50, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: euler, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  custom_sigmas: [],
  cache_mode: "cache-dit",
  cache_option: "",
  cache: enabled (threshold=1, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 42,
  upscale_repeats: 1,
  upscale_tile_size: 128,
}
[DEBUG] stable-diffusion.cpp:169  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:190  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:248  - loading model from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors', prefix = ''
[INFO ] stable-diffusion.cpp:271  - loading clip_l from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_l.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors', prefix = 'text_encoders.clip_l.transformer.'
[INFO ] stable-diffusion.cpp:279  - loading clip_g from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_g.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors', prefix = 'text_encoders.clip_g.transformer.'
[INFO ] stable-diffusion.cpp:295  - loading t5xxl from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors', prefix = 'text_encoders.t5xxl.transformer.'
[INFO ] stable-diffusion.cpp:316  - loading vae from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:332  - Version: SD3.x
[INFO ] stable-diffusion.cpp:360  - Weight type stat:                      f16: 1597 |    bf16: 244
[INFO ] stable-diffusion.cpp:361  - Conditioner weight type stat:          f16: 932
[INFO ] stable-diffusion.cpp:362  - Diffusion model weight type stat:      f16: 665
[INFO ] stable-diffusion.cpp:363  - VAE weight type stat:                 bf16: 244
[DEBUG] stable-diffusion.cpp:365  - ggml tensor size = 400 bytes
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[INFO ] mmdit.hpp:690  - MMDiT layers: 24 (including 13 MMDiT-x layers)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  235.06 MB(VRAM) (196 tensors)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  1329.29 MB(VRAM) (517 tensors)
[DEBUG] ggml_extend.hpp:1921 - t5 params backend buffer size =  9083.77 MB(VRAM) (219 tensors)
[DEBUG] ggml_extend.hpp:1921 - mmdit params backend buffer size =  5172.83 MB(VRAM) (665 tensors)
[INFO ] stable-diffusion.cpp:616  - Using Conv2d direct in the vae model
[DEBUG] ggml_extend.hpp:1921 - vae params backend buffer size =  94.57 MB(VRAM) (138 tensors)
[DEBUG] stable-diffusion.cpp:741  - loading weights
[DEBUG] model.cpp:1351 - using 15 threads for model loading
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================>                               | 665/1841 - 426.28it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_l.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=======================>                          | 861/1841 - 345.51it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_g.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=====================================>            | 1378/1841 - 465.70it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |===========================================>      | 1597/1841 - 322.89it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================================================| 1841/1841 - 277.30it/s
[INFO ] model.cpp:1599 - loading tensors completed, taking 6.66s (process: 0.00s, read: 1.95s, memcpy: 0.00s, convert: 0.04s, copy_to_backend: 0.61s)
[DEBUG] stable-diffusion.cpp:776  - finished loaded file
[INFO ] stable-diffusion.cpp:848  - total params memory size = 15915.53MB (VRAM 15915.53MB, RAM 0.00MB): text_encoders 10648.12MB(VRAM), diffusion_model 5172.83MB(VRAM), vae 94.57MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
[INFO ] stable-diffusion.cpp:917  - running in FLOW mode
[DEBUG] stable-diffusion.cpp:3460 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3494 - sampling using Euler method
[INFO ] denoiser.hpp:403  - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3621 - TXT2IMG
[DEBUG] conditioner.hpp:810  - parse 'a cat' to [['a cat', 1], ]
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1999 ms
[DEBUG] conditioner.hpp:810  - parse '' to [['', 1], ]
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1964 ms
[INFO ] stable-diffusion.cpp:3238 - get_learned_condition completed, taking 3969 ms
[INFO ] stable-diffusion.cpp:3349 - generating image: 1/1 - seed 42
[INFO ] stable-diffusion.cpp:1742 - CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
[DEBUG] ggml_extend.hpp:1733 - mmdit compute buffer size: 181.64 MB(VRAM)
  |==================================================| 20/20 - 6.78s/it
[INFO ] stable-diffusion.cpp:2219 - CacheDIT skipped 4/20 steps (1.25x estimated speedup), accum_diff: 0.2213
[INFO ] stable-diffusion.cpp:3391 - sampling completed, taking 135.77s
[INFO ] stable-diffusion.cpp:3402 - generating 1 latent images completed, taking 137.57s
[INFO ] stable-diffusion.cpp:3405 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1733 - vae compute buffer size: 704.25 MB(VRAM)
[DEBUG] stable-diffusion.cpp:2672 - computing vae decode graph completed, taking 8.12s
[INFO ] stable-diffusion.cpp:3415 - latent 1 decoded, taking 8.12s
[INFO ] stable-diffusion.cpp:3419 - decode_first_stage completed, taking 8.12s
[INFO ] stable-diffusion.cpp:3729 - generate_image completed in 149.67s
[INFO ] main.cpp:421  - save result image 0 to '.\cat_preset_fast.png' (success)

Preset MODE: medium

Preset MODE: medium
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 885e62e
[DEBUG] main.cpp:501  - System Info:
    SSE3 = 1 |     AVX = 1 |     AVX2 = 1 |     AVX512 = 0 |     AVX512_VBMI = 0 |     AVX512_VNNI = 0 |     FMA = 1 |     NEON = 0 |     ARM_FMA = 0 |     F16C = 1 |     FP16_VA = 0 |     WASM_SIMD = 0 |     VSX = 0 |
[DEBUG] main.cpp:502  - SDCliParams {
  mode: img_gen,
  output_path: ".\cat_preset_medium.png",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false
}
[DEBUG] main.cpp:503  - SDContextParams {
  n_threads: 15,
  model_path: "E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors",
  clip_l_path: "E:\AI\models\sd3_0_3_5\clip_l.safetensors",
  clip_g_path: "E:\AI\models\sd3_0_3_5\clip_g.safetensors",
  clip_vision_path: "",
  t5xxl_path: "E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors",
  llm_path: "",
  llm_vision_path: "",
  diffusion_model_path: "",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "",
  photo_maker_path: "",
  rng_type: cuda,
  sampler_rng_type: NONE,
  flow_shift: INF
  offload_params_to_cpu: false,
  enable_mmap: true,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  diffusion_flash_attn: false,
  diffusion_conv_direct: true,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: auto,
  vae_tiling_params: { 0, 0, 0, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "a cat",
  negative_prompt: "",
  clip_skip: -1,
  width: 512,
  height: 512,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 4.50, img_cfg: 4.50, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: euler, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  custom_sigmas: [],
  cache_mode: "cache-dit",
  cache_option: "",
  cache: enabled (threshold=1, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 42,
  upscale_repeats: 1,
  upscale_tile_size: 128,
}
[DEBUG] stable-diffusion.cpp:169  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:190  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:248  - loading model from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors', prefix = ''
[INFO ] stable-diffusion.cpp:271  - loading clip_l from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_l.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors', prefix = 'text_encoders.clip_l.transformer.'
[INFO ] stable-diffusion.cpp:279  - loading clip_g from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_g.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors', prefix = 'text_encoders.clip_g.transformer.'
[INFO ] stable-diffusion.cpp:295  - loading t5xxl from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors', prefix = 'text_encoders.t5xxl.transformer.'
[INFO ] stable-diffusion.cpp:316  - loading vae from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:332  - Version: SD3.x
[INFO ] stable-diffusion.cpp:360  - Weight type stat:                      f16: 1597 |    bf16: 244
[INFO ] stable-diffusion.cpp:361  - Conditioner weight type stat:          f16: 932
[INFO ] stable-diffusion.cpp:362  - Diffusion model weight type stat:      f16: 665
[INFO ] stable-diffusion.cpp:363  - VAE weight type stat:                 bf16: 244
[DEBUG] stable-diffusion.cpp:365  - ggml tensor size = 400 bytes
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[INFO ] mmdit.hpp:690  - MMDiT layers: 24 (including 13 MMDiT-x layers)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  235.06 MB(VRAM) (196 tensors)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  1329.29 MB(VRAM) (517 tensors)
[DEBUG] ggml_extend.hpp:1921 - t5 params backend buffer size =  9083.77 MB(VRAM) (219 tensors)
[DEBUG] ggml_extend.hpp:1921 - mmdit params backend buffer size =  5172.83 MB(VRAM) (665 tensors)
[INFO ] stable-diffusion.cpp:616  - Using Conv2d direct in the vae model
[DEBUG] ggml_extend.hpp:1921 - vae params backend buffer size =  94.57 MB(VRAM) (138 tensors)
[DEBUG] stable-diffusion.cpp:741  - loading weights
[DEBUG] model.cpp:1351 - using 15 threads for model loading
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================>                               | 665/1841 - 418.24it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_l.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=======================>                          | 861/1841 - 339.91it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_g.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=====================================>            | 1378/1841 - 459.03it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |===========================================>      | 1597/1841 - 324.26it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================================================| 1841/1841 - 282.49it/s
[INFO ] model.cpp:1599 - loading tensors completed, taking 6.54s (process: 0.00s, read: 1.91s, memcpy: 0.00s, convert: 0.04s, copy_to_backend: 0.60s)
[DEBUG] stable-diffusion.cpp:776  - finished loaded file
[INFO ] stable-diffusion.cpp:848  - total params memory size = 15915.53MB (VRAM 15915.53MB, RAM 0.00MB): text_encoders 10648.12MB(VRAM), diffusion_model 5172.83MB(VRAM), vae 94.57MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
[INFO ] stable-diffusion.cpp:917  - running in FLOW mode
[DEBUG] stable-diffusion.cpp:3460 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3494 - sampling using Euler method
[INFO ] denoiser.hpp:403  - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3621 - TXT2IMG
[DEBUG] conditioner.hpp:810  - parse 'a cat' to [['a cat', 1], ]
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1807 ms
[DEBUG] conditioner.hpp:810  - parse '' to [['', 1], ]
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1858 ms
[INFO ] stable-diffusion.cpp:3238 - get_learned_condition completed, taking 3671 ms
[INFO ] stable-diffusion.cpp:3349 - generating image: 1/1 - seed 42
[INFO ] stable-diffusion.cpp:1742 - CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
[DEBUG] ggml_extend.hpp:1733 - mmdit compute buffer size: 181.64 MB(VRAM)
  |==================================================| 20/20 - 6.65s/it
[INFO ] stable-diffusion.cpp:2219 - CacheDIT skipped 4/20 steps (1.25x estimated speedup), accum_diff: 0.2213
[INFO ] stable-diffusion.cpp:3391 - sampling completed, taking 133.08s
[INFO ] stable-diffusion.cpp:3402 - generating 1 latent images completed, taking 134.85s
[INFO ] stable-diffusion.cpp:3405 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1733 - vae compute buffer size: 704.25 MB(VRAM)
[DEBUG] stable-diffusion.cpp:2672 - computing vae decode graph completed, taking 8.06s
[INFO ] stable-diffusion.cpp:3415 - latent 1 decoded, taking 8.06s
[INFO ] stable-diffusion.cpp:3419 - decode_first_stage completed, taking 8.06s
[INFO ] stable-diffusion.cpp:3729 - generate_image completed in 146.60s
[INFO ] main.cpp:421  - save result image 0 to '.\cat_preset_medium.png' (success)

Preset MODE: slow

Preset MODE: slow
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 885e62e
[DEBUG] main.cpp:501  - System Info:
    SSE3 = 1 |     AVX = 1 |     AVX2 = 1 |     AVX512 = 0 |     AVX512_VBMI = 0 |     AVX512_VNNI = 0 |     FMA = 1 |     NEON = 0 |     ARM_FMA = 0 |     F16C = 1 |     FP16_VA = 0 |     WASM_SIMD = 0 |     VSX = 0 |
[DEBUG] main.cpp:502  - SDCliParams {
  mode: img_gen,
  output_path: ".\cat_preset_slow.png",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false
}
[DEBUG] main.cpp:503  - SDContextParams {
  n_threads: 15,
  model_path: "E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors",
  clip_l_path: "E:\AI\models\sd3_0_3_5\clip_l.safetensors",
  clip_g_path: "E:\AI\models\sd3_0_3_5\clip_g.safetensors",
  clip_vision_path: "",
  t5xxl_path: "E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors",
  llm_path: "",
  llm_vision_path: "",
  diffusion_model_path: "",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "",
  photo_maker_path: "",
  rng_type: cuda,
  sampler_rng_type: NONE,
  flow_shift: INF
  offload_params_to_cpu: false,
  enable_mmap: true,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  diffusion_flash_attn: false,
  diffusion_conv_direct: true,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: auto,
  vae_tiling_params: { 0, 0, 0, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "a cat",
  negative_prompt: "",
  clip_skip: -1,
  width: 512,
  height: 512,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 4.50, img_cfg: 4.50, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: euler, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: 0.00, shifted_timestep: 0),
  custom_sigmas: [],
  cache_mode: "cache-dit",
  cache_option: "",
  cache: enabled (threshold=1, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 42,
  upscale_repeats: 1,
  upscale_tile_size: 128,
}
[DEBUG] stable-diffusion.cpp:169  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:190  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:248  - loading model from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors', prefix = ''
[INFO ] stable-diffusion.cpp:271  - loading clip_l from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_l.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_l.safetensors', prefix = 'text_encoders.clip_l.transformer.'
[INFO ] stable-diffusion.cpp:279  - loading clip_g from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\clip_g.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\clip_g.safetensors', prefix = 'text_encoders.clip_g.transformer.'
[INFO ] stable-diffusion.cpp:295  - loading t5xxl from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors', prefix = 'text_encoders.t5xxl.transformer.'
[INFO ] stable-diffusion.cpp:316  - loading vae from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors'
[INFO ] model.cpp:373  - load E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:332  - Version: SD3.x
[INFO ] stable-diffusion.cpp:360  - Weight type stat:                      f16: 1597 |    bf16: 244
[INFO ] stable-diffusion.cpp:361  - Conditioner weight type stat:          f16: 932
[INFO ] stable-diffusion.cpp:362  - Diffusion model weight type stat:      f16: 665
[INFO ] stable-diffusion.cpp:363  - VAE weight type stat:                 bf16: 244
[DEBUG] stable-diffusion.cpp:365  - ggml tensor size = 400 bytes
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[DEBUG] clip.hpp:160  - vocab size: 49408
[DEBUG] clip.hpp:171  - trigger word img already in vocab
[INFO ] mmdit.hpp:690  - MMDiT layers: 24 (including 13 MMDiT-x layers)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  235.06 MB(VRAM) (196 tensors)
[DEBUG] ggml_extend.hpp:1921 - clip params backend buffer size =  1329.29 MB(VRAM) (517 tensors)
[DEBUG] ggml_extend.hpp:1921 - t5 params backend buffer size =  9083.77 MB(VRAM) (219 tensors)
[DEBUG] ggml_extend.hpp:1921 - mmdit params backend buffer size =  5172.83 MB(VRAM) (665 tensors)
[INFO ] stable-diffusion.cpp:616  - Using Conv2d direct in the vae model
[DEBUG] ggml_extend.hpp:1921 - vae params backend buffer size =  94.57 MB(VRAM) (138 tensors)
[DEBUG] stable-diffusion.cpp:741  - loading weights
[DEBUG] model.cpp:1351 - using 15 threads for model loading
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\sd3.5m_turbo.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================>                               | 665/1841 - 431.54it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_l.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=======================>                          | 861/1841 - 346.62it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\clip_g.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |=====================================>            | 1378/1841 - 469.67it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\t5xxl_fp16.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |===========================================>      | 1597/1841 - 328.06it/s
[DEBUG] model.cpp:1373 - loading tensors from E:\AI\models\sd3_0_3_5\vae4sd3_diffusion_pytorch_model.safetensors
[DEBUG] model.cpp:1395 - using mmap for I/O
  |==================================================| 1841/1841 - 284.81it/s
[INFO ] model.cpp:1599 - loading tensors completed, taking 6.49s (process: 0.00s, read: 1.90s, memcpy: 0.00s, convert: 0.04s, copy_to_backend: 0.62s)
[DEBUG] stable-diffusion.cpp:776  - finished loaded file
[INFO ] stable-diffusion.cpp:848  - total params memory size = 15915.53MB (VRAM 15915.53MB, RAM 0.00MB): text_encoders 10648.12MB(VRAM), diffusion_model 5172.83MB(VRAM), vae 94.57MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
[INFO ] stable-diffusion.cpp:917  - running in FLOW mode
[DEBUG] stable-diffusion.cpp:3460 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3494 - sampling using Euler method
[INFO ] denoiser.hpp:403  - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3621 - TXT2IMG
[DEBUG] conditioner.hpp:810  - parse 'a cat' to [['a cat', 1], ]
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:401  - token a cat
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1738 ms
[DEBUG] conditioner.hpp:810  - parse '' to [['', 1], ]
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] clip.hpp:304  - token length: 77
[DEBUG] t5.hpp:402  - token length: 77
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 1.40 MB(VRAM)
[DEBUG] clip.hpp:764  - identity projection
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - clip compute buffer size: 2.33 MB(VRAM)
[DEBUG] ggml_extend.hpp:1733 - t5 compute buffer size: 11.94 MB(VRAM)
[DEBUG] conditioner.hpp:1077 - computing condition graph completed, taking 1787 ms
[INFO ] stable-diffusion.cpp:3238 - get_learned_condition completed, taking 3531 ms
[INFO ] stable-diffusion.cpp:3349 - generating image: 1/1 - seed 42
[INFO ] stable-diffusion.cpp:1742 - CacheDIT enabled - mode: DBCache+TaylorSeer, Fn: 8, Bn: 0, threshold: 0.080, warmup: 8
[DEBUG] ggml_extend.hpp:1733 - mmdit compute buffer size: 181.64 MB(VRAM)
  |==================================================| 20/20 - 6.71s/it
[INFO ] stable-diffusion.cpp:2219 - CacheDIT skipped 4/20 steps (1.25x estimated speedup), accum_diff: 0.2213
[INFO ] stable-diffusion.cpp:3391 - sampling completed, taking 134.16s
[INFO ] stable-diffusion.cpp:3402 - generating 1 latent images completed, taking 135.80s
[INFO ] stable-diffusion.cpp:3405 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1733 - vae compute buffer size: 704.25 MB(VRAM)
[DEBUG] stable-diffusion.cpp:2672 - computing vae decode graph completed, taking 7.98s
[INFO ] stable-diffusion.cpp:3415 - latent 1 decoded, taking 7.99s
[INFO ] stable-diffusion.cpp:3419 - decode_first_stage completed, taking 7.99s
[INFO ] stable-diffusion.cpp:3729 - generate_image completed in 147.33s
[INFO ] main.cpp:421  - save result image 0 to '.\cat_preset_slow.png' (success)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Is it expected that the --cache-preset option doesn't work? #1188

Uh oh!

{{title}}

Uh oh!

My environment:

Results and Logs

Summary:

All logs:

Replies: 0 comments

Select a reply

Uh oh!

Is it expected that the --cache-preset option doesn't work? #1188

Uh oh!

masamaru-san Jan 12, 2026

❓Question:

My environment:

Results and Logs

Summary:

All logs:

Replies: 0 comments

masamaru-san
Jan 12, 2026