[Bug] "CLIP: Using CPU backend" and then crash

[AndriiParf]

Git commit

61659ef

Operating System & Version

Windows 11 24H2

GGML backends

Vulkan

Command-line arguments used

sd-cli.exe --diffusion-model E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf --vae E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors --llm E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-8B-Q6_K.gguf -r example.png -p "change the background to park" -W 960 -H 960 --cfg-scale 4.0 --steps 4 -v --offload-to-cpu --diffusion-fa -t 12 --sampling-method dpm++2mv2 --clip-on-cpu --vae-tiling --vae-tile-size 64x64

Steps to reproduce

Checkout commit 61659ef.
Run the generation command with Flux 2 Klein 4B.
The application initializes the system info, loads the arguments, detects the Vulkan device, and prints weight statistics.
Immediately after printing CLIP: Using CPU backend, the process terminates (or hangs) without any further error message or assertion failure.

What you expected to happen

The application should continue to load the LLM vocab/merges and then proceed to load model tensors, as it did in previous versions.

What actually happened

The execution stops abruptly at initialization.

Logs / error messages / stack trace

sd-cli.exe --diffusion-model E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf --vae E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors --llm E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-8B-Q6_K.gguf -r example.png -p "change the background to park" -W 960 -H 960 --cfg-scale 4.0 --steps 4 -v --offload-to-cpu --diffusion-fa -t 12 --sampling-method dpm++2mv2 --clip-on-cpu --vae-tiling --vae-tile-size 64x64
[DEBUG] main.cpp:500 - version: stable-diffusion.cpp version unknown, commit 61659ef
[DEBUG] main.cpp:501 - System Info:
SSE3 = 1 | AVX = 1 | AVX2 = 1 | AVX512 = 1 | 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: "output.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: 12,
model_path: "",
clip_l_path: "",
clip_g_path: "",
clip_vision_path: "",
t5xxl_path: "",
llm_path: "E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-8B-Q6_K.gguf",
llm_vision_path: "",
diffusion_model_path: "E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf",
high_noise_diffusion_model_path: "",
vae_path: "E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.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: true,
enable_mmap: false,
control_net_cpu: false,
clip_on_cpu: true,
vae_on_cpu: true,
diffusion_flash_attn: true,
diffusion_conv_direct: false,
vae_conv_direct: false,
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: { 1, 64, 64, 0.5, 0, 0 },
force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504 - SDGenerationParams {
loras: "{
}",
high_noise_loras: "{
}",
prompt: "change the background to park",
negative_prompt: "",
clip_skip: -1,
width: 960,
height: 960,
batch_count: 1,
init_image_path: "",
end_image_path: "",
mask_image_path: "",
control_image_path: "",
ref_image_paths: ["example.png"],
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: 4.00, img_cfg: 4.00, 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: dpm++2mv2, sample_steps: 4, 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_option: "",
cache: disabled (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:171 - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75 - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75 - ggml_vulkan: 0 = Radeon RX 580 Series (AMD proprietary driver) | uma: 0 | fp16: 0 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:192 - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:257 - loading diffusion model from 'E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf'
[INFO ] model.cpp:370 - load E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf using gguf format
[DEBUG] model.cpp:412 - init from 'E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:304 - loading llm from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-8B-Q6_K.gguf'
[INFO ] model.cpp:370 - load E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-8B-Q6_K.gguf using gguf format
[DEBUG] model.cpp:412 - init from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-8B-Q6_K.gguf'
[INFO ] stable-diffusion.cpp:318 - loading vae from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors'
[INFO ] model.cpp:373 - load E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors using safetensors format
[DEBUG] model.cpp:503 - init from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:334 - Version: Flux.2 klein
[INFO ] stable-diffusion.cpp:362 - Weight type stat: f32: 393 | q8_0: 80 | q6_K: 253 | bf16: 69
[INFO ] stable-diffusion.cpp:363 - Conditioner weight type stat: f32: 145 | q6_K: 253
[INFO ] stable-diffusion.cpp:364 - Diffusion model weight type stat: q8_0: 80 | bf16: 69
[INFO ] stable-diffusion.cpp:365 - VAE weight type stat: f32: 248
[DEBUG] stable-diffusion.cpp:367 - ggml tensor size = 400 bytes
[INFO ] stable-diffusion.cpp:426 - CLIP: Using CPU backend

Additional context / environment details

GPU: Radeon RX 580 8GB (Vulkan)
Models: Flux 2 Klein 4b Q8_0, Qwen3-8B-Q6_K
In commit 9565c7f, progress continues, but an error also appears (a different error, not this one).

[stduhpf]

I think that's because Flux 2 Klein 4b requires Qwen3-4B, not Qwen3-8B. The lack of error message to indicate that the it's not the right text encoder is definitely an issue in this case.

[AndriiParf]

Oops, thank you, and sorry for bothering you.

[AndriiParf]

I downloaded Qwen3 4b, but the problem remains. I checked the file paths and there are no problems.

Logs:

sd-cli.exe --diffusion-model  E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf --vae E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors  --llm E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf -r example.png -p "change the background to park" -W 960 -H 960 --cfg-scale 4.0 --steps 4 -v --offload-to-cpu --diffusion-fa -t 12 --sampling-method dpm++2mv2 --clip-on-cpu --vae-tiling --vae-tile-size 64x64
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 61659ef
[DEBUG] main.cpp:501  - System Info:
    SSE3 = 1 |     AVX = 1 |     AVX2 = 1 |     AVX512 = 1 |     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: "output.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: 12,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.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: true,
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: true,
  vae_on_cpu: false,
  diffusion_flash_attn: true,
  diffusion_conv_direct: false,
  vae_conv_direct: false,
  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: { 1, 64, 64, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "change the background to park",
  negative_prompt: "",
  clip_skip: -1,
  width: 960,
  height: 960,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: ["example.png"],
  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: 4.00, img_cfg: 4.00, 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: dpm++2mv2, sample_steps: 4, 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_option: "",
  cache: disabled (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:171  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = Radeon RX 580 Series (AMD proprietary driver) | uma: 0 | fp16: 0 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:192  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:257  - loading diffusion model from 'E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf'
[INFO ] model.cpp:370  - load E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf using gguf format
[DEBUG] model.cpp:412  - init from 'E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:304  - loading llm from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf'
[INFO ] model.cpp:370  - load E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf using gguf format
[DEBUG] model.cpp:412  - init from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:318  - loading vae from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors'
[INFO ] model.cpp:373  - load E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:334  - Version: Flux.2 klein
[INFO ] stable-diffusion.cpp:362  - Weight type stat:                      f32: 393  |    q8_0: 333  |    bf16: 69
[INFO ] stable-diffusion.cpp:363  - Conditioner weight type stat:          f32: 145  |    q8_0: 253
[INFO ] stable-diffusion.cpp:364  - Diffusion model weight type stat:     q8_0: 80   |    bf16: 69
[INFO ] stable-diffusion.cpp:365  - VAE weight type stat:                  f32: 248
[DEBUG] stable-diffusion.cpp:367  - ggml tensor size = 400 bytes
[INFO ] stable-diffusion.cpp:426  - CLIP: Using CPU backend

[AndriiParf]

There is no problem in commit 9565c7f.

Logs:

sd-cli.exe --diffusion-model  E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf --vae E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors  --llm E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf -p "park, sunny, trees" -W 512 -H 512 --cfg-scale 4.0 --steps 4 -v --offload-to-cpu --diffusion-fa -t 12 --sampling-method dpm++2mv2 --clip-on-cpu --vae-tiling --vae-tile-size 64x64
[DEBUG] main.cpp:500  - version: stable-diffusion.cpp version unknown, commit 9565c7f
[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: "output.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: 12,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  vae_path: "E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.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: true,
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: true,
  vae_on_cpu: false,
  diffusion_flash_attn: true,
  diffusion_conv_direct: false,
  vae_conv_direct: false,
  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: { 1, 64, 64, 0.5, 0, 0 },
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "park, sunny, trees",
  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: 4.00, img_cfg: 4.00, 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: dpm++2mv2, sample_steps: 4, 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_option: "",
  cache: disabled (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:171  - Using Vulkan backend
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: Found 1 Vulkan devices:
[DEBUG] ggml_extend.hpp:75   - ggml_vulkan: 0 = Radeon RX 580 Series (AMD proprietary driver) | uma: 0 | fp16: 0 | bf16: 0 | warp size: 64 | shared memory: 32768 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:192  - Vulkan: Using device 0
[INFO ] stable-diffusion.cpp:257  - loading diffusion model from 'E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf'
[INFO ] model.cpp:370  - load E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf using gguf format
[DEBUG] model.cpp:412  - init from 'E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:304  - loading llm from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf'
[INFO ] model.cpp:370  - load E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf using gguf format
[DEBUG] model.cpp:412  - init from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:318  - loading vae from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors'
[INFO ] model.cpp:373  - load E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors using safetensors format
[DEBUG] model.cpp:503  - init from 'E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:334  - Version: Flux.2 klein
[INFO ] stable-diffusion.cpp:362  - Weight type stat:                      f32: 393  |    q8_0: 333  |    bf16: 69
[INFO ] stable-diffusion.cpp:363  - Conditioner weight type stat:          f32: 145  |    q8_0: 253
[INFO ] stable-diffusion.cpp:364  - Diffusion model weight type stat:     q8_0: 80   |    bf16: 69
[INFO ] stable-diffusion.cpp:365  - VAE weight type stat:                  f32: 248
[DEBUG] stable-diffusion.cpp:367  - ggml tensor size = 400 bytes
[INFO ] stable-diffusion.cpp:426  - CLIP: Using CPU backend
[DEBUG] llm.hpp:285  - merges size 151387
[DEBUG] llm.hpp:317  - vocab size: 151669
[DEBUG] llm.hpp:1139 - llm: num_layers = 36, vocab_size = 151936, hidden_size = 2560, intermediate_size = 9728
[INFO ] flux.hpp:1364 - flux: depth = 5, depth_single_blocks = 20, guidance_embed = false, context_in_dim = 7680, hidden_size = 3072, num_heads = 24
[INFO ] stable-diffusion.cpp:572  - Using flash attention in the diffusion model
[DEBUG] ggml_extend.hpp:1922 - qwen3 params backend buffer size =  4076.43 MB(RAM) (398 tensors)
[DEBUG] ggml_extend.hpp:1922 - flux params backend buffer size =  4101.40 MB(RAM) (149 tensors)
[DEBUG] ggml_extend.hpp:1922 - vae params backend buffer size =  94.72 MB(RAM) (140 tensors)
[DEBUG] stable-diffusion.cpp:751  - loading weights
[DEBUG] model.cpp:1370 - using 12 threads for model loading
[DEBUG] model.cpp:1392 - loading tensors from E:\GenerateImage\Models\FLUX\Realistic\flux-2-klein-4b-Q8_0.gguf
  |=========>                                        | 149/795 - 144.38it/s
[DEBUG] model.cpp:1392 - loading tensors from E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\Qwen3-4B-Q8_0.gguf
  |==================================>               | 547/795 - 265.79it/s
[DEBUG] model.cpp:1392 - loading tensors from E:\GenerateImage\SDXLVaeAndClips\FLUX2Klein\ae.safetensors
  |==================================================| 795/795 - 350.84it/s
[INFO ] model.cpp:1618 - loading tensors completed, taking 2.27s (process: 0.00s, read: 1.96s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s)
[DEBUG] stable-diffusion.cpp:786  - finished loaded file
[INFO ] stable-diffusion.cpp:859  - total params memory size = 8272.55MB (VRAM 4196.12MB, RAM 4076.43MB): text_encoders 4076.43MB(RAM), diffusion_model 4101.40MB(VRAM), vae 94.72MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(RAM)
[INFO ] stable-diffusion.cpp:938  - running in Flux2 FLOW mode
[DEBUG] stable-diffusion.cpp:3471 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3505 - sampling using modified DPM++ (2M) method
[DEBUG] denoiser.hpp:703  - Flux2FlowDenoiser: set shift to 2.031
[INFO ] denoiser.hpp:403  - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3632 - TXT2IMG
[DEBUG] conditioner.hpp:1679 - parse '<|im_start|>user
park, sunny, trees<|im_end|>
<|im_start|>assistant
<think>

</think>

' to [['<|im_start|>user
', 1], ['park, sunny, trees', 1], ['<|im_end|>
<|im_start|>assistant
<think>

</think>

', 1], ]
[DEBUG] llm.hpp:259  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] llm.hpp:259  - split prompt "park, sunny, trees" to tokens ["park", ",", "Ġsunny", ",", "Ġtrees", ]
[DEBUG] llm.hpp:259  - split prompt "<|im_end|>
<|im_start|>assistant
<think>

</think>

" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", "<think>", "ĊĊ", "</think>", "ĊĊ", ]
[DEBUG] llm.hpp:203  - token length: 512
[DEBUG] ggml_extend.hpp:1734 - qwen3 compute buffer size: 74.00 MB(RAM)
[DEBUG] conditioner.hpp:1923 - computing condition graph completed, taking 9856 ms
[DEBUG] conditioner.hpp:1679 - parse '<|im_start|>user
<|im_end|>
<|im_start|>assistant
<think>

</think>

' to [['<|im_start|>user
', 1], ['<|im_end|>
<|im_start|>assistant
<think>

</think>

', 1], ]
[DEBUG] llm.hpp:259  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] llm.hpp:259  - split prompt "<|im_end|>
<|im_start|>assistant
<think>

</think>

" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", "<think>", "ĊĊ", "</think>", "ĊĊ", ]
[DEBUG] llm.hpp:203  - token length: 512
[DEBUG] ggml_extend.hpp:1734 - qwen3 compute buffer size: 74.00 MB(RAM)
[DEBUG] conditioner.hpp:1923 - computing condition graph completed, taking 9834 ms
[INFO ] stable-diffusion.cpp:3249 - get_learned_condition completed, taking 19693 ms
[INFO ] stable-diffusion.cpp:3360 - generating image: 1/1 - seed 42
[INFO ] ggml_extend.hpp:1834 - flux offload params (4101.40 MB, 149 tensors) to runtime backend (Vulkan0), taking 0.97s
[DEBUG] ggml_extend.hpp:1734 - flux compute buffer size: 377.00 MB(VRAM)
  |==================================================| 4/4 - 19.47s/it
[INFO ] stable-diffusion.cpp:3402 - sampling completed, taking 78.03s
[INFO ] stable-diffusion.cpp:3413 - generating 1 latent images completed, taking 78.35s
[INFO ] stable-diffusion.cpp:3416 - decoding 1 latents
[DEBUG] stable-diffusion.cpp:2657 - VAE Tile size: 32x32
[DEBUG] ggml_extend.hpp:849  - num tiles : 1, 1
[DEBUG] ggml_extend.hpp:850  - optimal overlap : 0.000000, 0.000000 (targeting 0.500000)
[DEBUG] ggml_extend.hpp:881  - tile work buffer size: 3.50 MB
[DEBUG] ggml_extend.hpp:894  - processing 1 tiles
[INFO ] ggml_extend.hpp:1834 - vae offload params ( 94.72 MB, 140 tensors) to runtime backend (Vulkan0), taking 0.03s
[DEBUG] ggml_extend.hpp:1734 - vae compute buffer size: 1984.50 MB(VRAM)
  |==================================================| 1/1 - 5.28s/it
[DEBUG] stable-diffusion.cpp:2683 - computing vae decode graph completed, taking 5.30s
[INFO ] stable-diffusion.cpp:3426 - latent 1 decoded, taking 5.30s
[INFO ] stable-diffusion.cpp:3430 - decode_first_stage completed, taking 5.30s
[INFO ] stable-diffusion.cpp:3740 - generate_image completed in 103.35s
[INFO ] main.cpp:421  - save result image 0 to 'output.png' (success)