[Bug] master-493-65891d7 broken

[Korsar13]

Git commit

65891d7

Operating System & Version

Windows 11

GGML backends

CUDA

Command-line arguments used

sd-cli.exe -v --mmap --steps 5 --diffusion-model ..\models\diff\z_image_turbo-Q4_K.gguf --llm ..\models\llm\Qwen3-4B-Instruct-2507-Q8_0.gguf --vae ..\models\vae\pig_flux_vae_fp32-f16.gguf --diffusion-fa --rng cpu --sampler-rng cpu --clip-on-cpu --vae-tiling --cfg-scale 1.0 -H 512 -W 512 --output test.png -p "A paper dragon"

Steps to reproduce

run sd-cli.exe

What you expected to happen

.png generated (as for commit f957fa3)

What actually happened

crashed

Logs / error messages / stack trace

d:\bin\sdcpp\cuda>sd-cli.exe -v --mmap --steps 5 --diffusion-model ..\models\diff\z_image_turbo-Q4_K.gguf --llm ..\models\llm\Qwen3-4B-Instruct-2507-Q8_0.gguf --vae ..\models\vae\pig_flux_vae_fp32-f16.gguf --diffusion-fa --rng cpu --sampler-rng cpu --clip-on-cpu --vae-tiling --cfg-scale 1.0 -H 512 -W 512 --output test.png -p "A paper dragon"
[DEBUG] main.cpp:500 - version: stable-diffusion.cpp version unknown, commit 65891d7
[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: "test.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: 6,
model_path: "",
clip_l_path: "",
clip_g_path: "",
clip_vision_path: "",
t5xxl_path: "",
llm_path: "..\models\llm\Qwen3-4B-Instruct-2507-Q8_0.gguf",
llm_vision_path: "",
diffusion_model_path: "..\models\diff\z_image_turbo-Q4_K.gguf",
high_noise_diffusion_model_path: "",
vae_path: "..\models\vae\pig_flux_vae_fp32-f16.gguf",
taesd_path: "",
esrgan_path: "",
control_net_path: "",
embedding_dir: "",
embeddings: {
}
wtype: NONE,
tensor_type_rules: "",
lora_model_dir: ".",
photo_maker_path: "",
rng_type: cpu,
sampler_rng_type: cpu,
flow_shift: INF
offload_params_to_cpu: false,
enable_mmap: true,
control_net_cpu: false,
clip_on_cpu: true,
vae_on_cpu: false,
flash_attn: 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, 0, 0, 0.5, 0, 0 },
force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504 - SDGenerationParams {
loras: "{
}",
high_noise_loras: "{
}",
prompt: "A paper dragon",
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: 1.00, img_cfg: 1.00, 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: 5, 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:166 - Using CUDA backend
[INFO ] ggml_extend.hpp:78 - ggml_cuda_init: found 1 CUDA devices:
[INFO ] ggml_extend.hpp:78 - Device 0: NVIDIA GeForce GTX 1070, compute capability 6.1, VMM: yes
[INFO ] stable-diffusion.cpp:260 - loading diffusion model from '..\models\diff\z_image_turbo-Q4_K.gguf'
[INFO ] model.cpp:370 - load ..\models\diff\z_image_turbo-Q4_K.gguf using gguf format
[DEBUG] model.cpp:416 - init from '..\models\diff\z_image_turbo-Q4_K.gguf'
[INFO ] stable-diffusion.cpp:307 - loading llm from '..\models\llm\Qwen3-4B-Instruct-2507-Q8_0.gguf'
[INFO ] model.cpp:370 - load ..\models\llm\Qwen3-4B-Instruct-2507-Q8_0.gguf using gguf format
[DEBUG] model.cpp:416 - init from '..\models\llm\Qwen3-4B-Instruct-2507-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:321 - loading vae from '..\models\vae\pig_flux_vae_fp32-f16.gguf'
[INFO ] model.cpp:370 - load ..\models\vae\pig_flux_vae_fp32-f16.gguf using gguf format
[DEBUG] model.cpp:416 - init from '..\models\vae\pig_flux_vae_fp32-f16.gguf'
[INFO ] stable-diffusion.cpp:337 - Version: Z-Image
[INFO ] stable-diffusion.cpp:365 - Weight type stat: f32: 570 | f16: 70 | q8_0: 275 | q4_K: 180
[INFO ] stable-diffusion.cpp:366 - Conditioner weight type stat: f32: 145 | q8_0: 253
[INFO ] stable-diffusion.cpp:367 - Diffusion model weight type stat: f32: 251 | q8_0: 22 | q4_K: 180
[INFO ] stable-diffusion.cpp:368 - VAE weight type stat: f32: 174 | f16: 70
[DEBUG] stable-diffusion.cpp:370 - ggml tensor size = 400 bytes
[INFO ] stable-diffusion.cpp:429 - CLIP: Using CPU backend

d:\bin\sdcpp\cuda>

Additional context / environment details

i7-8700, GTX 1070 8GB, RAM 64 GB

[wbruna]

Possibly related to #1236 .

Does it work on older builds? What about without --mmap?

[Korsar13]

previous release [master-492-f957fa3] works as expected

[stduhpf]

That's very odd, the changes from 65891d7 should not affect anything but Qwen-image. Is it with the release binaries or built locally?

[Korsar13]

binaries from released archive

[stduhpf]

Does it also crash without --clip-on-cpu?

[anon0730]

Observing the same on the sd-master-65891d7-bin-win-cuda12-x64 build, not using --clip on-cpu myself. Neither Flux-Klein nor Z-Image runs. (to clarify: Those were all I tried and all I have downloaded right now so other model behavior unknown.)
No issues with either on 492-f957fa3.
System: Windows 10, Ryzen 5 3600, 64GB RAM, RTX 2070 Super 8GB

edit: might as well post my console output too.

W:\AI\sdcpp>.\sd-cli.exe -v --llm .\Qwen3-8B-Q5_K_M.gguf --diffusion-model .\flux-2-klein-9b-Q4_K_S.gguf --vae .\ae-klein.safetensors --lora-model-dir "W:\AI\sdcpp\lora" --offload-to-cpu --diffusion-fa --cfg-scale 1.0 --sampling-method euler --steps 6 -H 1024 -W 1024 --seed -1 -p "A woman holding a watermelon and smiling." -o test.png
[DEBUG] main.cpp:500 - version: stable-diffusion.cpp version unknown, commit 65891d7
[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: "test.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: 6,
model_path: "",
clip_l_path: "",
clip_g_path: "",
clip_vision_path: "",
t5xxl_path: "",
llm_path: ".\Qwen3-8B-Q5_K_M.gguf",
llm_vision_path: "",
diffusion_model_path: ".\flux-2-klein-9b-Q4_K_S.gguf",
high_noise_diffusion_model_path: "",
vae_path: ".\ae-klein.safetensors",
taesd_path: "",
esrgan_path: "",
control_net_path: "",
embedding_dir: "",
embeddings: {
}
wtype: NONE,
tensor_type_rules: "",
lora_model_dir: "W:\AI\sdcpp\lora",
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: false,
vae_on_cpu: false,
flash_attn: 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: { 0, 0, 0, 0.5, 0, 0 },
force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:504 - SDGenerationParams {
loras: "{
}",
high_noise_loras: "{
}",
prompt: "A woman holding a watermelon and smiling.",
negative_prompt: "",
clip_skip: -1,
width: 1024,
height: 1024,
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: 1.00, img_cfg: 1.00, 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: euler, sample_steps: 6, 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: 28192,
upscale_repeats: 1,
upscale_tile_size: 128,
}
[DEBUG] stable-diffusion.cpp:166 - Using CUDA backend
[INFO ] ggml_extend.hpp:78 - ggml_cuda_init: found 1 CUDA devices:
[INFO ] ggml_extend.hpp:78 - Device 0: NVIDIA GeForce RTX 2070 SUPER, compute capability 7.5, VMM: yes
[INFO ] stable-diffusion.cpp:260 - loading diffusion model from '.\flux-2-klein-9b-Q4_K_S.gguf'
[INFO ] model.cpp:370 - load .\flux-2-klein-9b-Q4_K_S.gguf using gguf format
[DEBUG] model.cpp:416 - init from '.\flux-2-klein-9b-Q4_K_S.gguf'
[INFO ] stable-diffusion.cpp:307 - loading llm from '.\Qwen3-8B-Q5_K_M.gguf'
[INFO ] model.cpp:370 - load .\Qwen3-8B-Q5_K_M.gguf using gguf format
[DEBUG] model.cpp:416 - init from '.\Qwen3-8B-Q5_K_M.gguf'
[INFO ] stable-diffusion.cpp:321 - loading vae from '.\ae-klein.safetensors'
[INFO ] model.cpp:373 - load .\ae-klein.safetensors using safetensors format
[DEBUG] model.cpp:507 - init from '.\ae-klein.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:337 - Version: Flux.2 klein
[INFO ] stable-diffusion.cpp:365 - Weight type stat: f32: 225 | q4_K: 88 | q5_K: 241 | q6_K: 36 | bf16: 257
[INFO ] stable-diffusion.cpp:366 - Conditioner weight type stat: f32: 145 | q5_K: 217 | q6_K: 36
[INFO ] stable-diffusion.cpp:367 - Diffusion model weight type stat: f32: 80 | q4_K: 88 | q5_K: 24 | bf16: 9
[INFO ] stable-diffusion.cpp:368 - VAE weight type stat: bf16: 248
[DEBUG] stable-diffusion.cpp:370 - ggml tensor size = 400 bytes
[DEBUG] llm.hpp:285 - merges size 151387
[DEBUG] llm.hpp:317 - vocab size: 151669
PS W:\AI\sdcpp>

[anon0730]

Update: Works on latest version after removing --offload-to-cpu.