i think something is broken for us cpu users [avx2]

[Amin456789]

hi

no matter what, it always crash and close the app itself after loading model or operating. i mean no matter img2img or txt or video
it worked well in a couple of weeks ago builds but now it is broken and no error is being shown
in -v debug it says= [debug] stable-diffusion.cpp:174 - using cpu backend then it has a window pop up sd.exe has stopped working
tnq

[Amin456789]

@LostRuins could u please take a look into it? i have no problem generation image with ur koboldcpp as it uses my avx2 very good but working with sd cpp doesnt work for me anymore

[wbruna]

Could you please give more details, so I can try to reproduce it here? Which version, model type, and parameters? Are you using a binary release, and on which platform?

[Amin456789]

hi, thank u for ur attention,
i use windows 11 and for the test i tested wan 2.1 gguf and the official parameters for txt and image generation. i use photon lcm model that i use for older version of sd cpp which works for image generation
even model conversaion doesnt work, i mean any operation doesnt work at all like before, i fairly know how to use sd cpp for a long time so i can tell something is wrong here unfurtunalty
yes i use binary release from this repo. the latest version

[wbruna]

I don't have Windows here, but the sd-master-abb115c-bin-win-avx2-x64.zip binary seems to work fine on Wine (I don't have the merged model, so I'm using the LCM LoRA):

wine ./sd.exe -p 'forestlora:lcm-lora-sdv1-5:1.0' --sampling-method lcm --model ./photon_v1.safetensors --lora-model-dir . --steps 6 --cfg-scale 1 -H 320

Perhaps the specific sd.exe you downloaded got corrupted?

[Amin456789]

thank u for ur answer, downloaded ur version. still no luck..

[wbruna]

Could you run that command with -v, and paste the result here (including the command itself)?

Also, what is the last version that do work for you?

[Amin456789]

this is the last version that works =
https://github.com/leejet/stable-diffusion.cpp/releases/tag/master-b1fc16b

after that everything else is not working.

the command that i work which works for the link above

PS C:\Users\Amin\Desktop\AI\Koboldcpp> ./sd.exe -v -m plcm.gguf -H 320 -W 320 -o ./pic.png --steps 4 -p "a hamster"
Option:
n_threads: 4
mode: txt2img
model_path: plcm.gguf
wtype: unspecified
clip_l_path:
clip_g_path:
t5xxl_path:
diffusion_model_path:
vae_path:
taesd_path:
esrgan_path:
controlnet_path:
embeddings_path:
stacked_id_embeddings_path:
input_id_images_path:
style ratio: 20.00
normalize input image : false
output_path: ./pic.png
init_img:
mask_img:
control_image:
ref_images_paths:
clip on cpu: false
controlnet cpu: false
vae decoder on cpu:false
diffusion flash attention:false
strength(control): 0.90
prompt: a hamster
negative_prompt:
min_cfg: 1.00
cfg_scale: 7.00
slg_scale: 0.00
guidance: 3.50
eta: 0.00
clip_skip: -1
width: 320
height: 320
sample_method: euler_a
schedule: default
sample_steps: 4
strength(img2img): 0.75
rng: cuda
seed: 42
batch_count: 1
vae_tiling: false
upscale_repeats: 1
chroma_use_dit_mask: true
chroma_use_t5_mask: false
chroma_t5_mask_pad: 1
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] stable-diffusion.cpp:190 - Using CPU backend
[INFO ] stable-diffusion.cpp:199 - loading model from 'plcm.gguf'
[INFO ] model.cpp:905 - load plcm.gguf using gguf format
[DEBUG] model.cpp:922 - init from 'plcm.gguf'
[INFO ] stable-diffusion.cpp:246 - Version: SD 1.x
[INFO ] stable-diffusion.cpp:279 - Weight type: q4_0
[INFO ] stable-diffusion.cpp:280 - Conditioner weight type: q4_0
[INFO ] stable-diffusion.cpp:281 - Diffusion model weight type: q4_0
[INFO ] stable-diffusion.cpp:282 - VAE weight type: q4_0
[DEBUG] stable-diffusion.cpp:284 - ggml tensor size = 400 bytes
[DEBUG] clip.hpp:171 - vocab size: 49408
[DEBUG] clip.hpp:182 - trigger word img already in vocab
[DEBUG] ggml_extend.hpp:1190 - clip params backend buffer size = 191.00 MB(RAM) (196 tensors)
[DEBUG] ggml_extend.hpp:1190 - unet params backend buffer size = 1272.85 MB(RAM) (686 tensors)
[DEBUG] ggml_extend.hpp:1190 - vae params backend buffer size = 94.47 MB(RAM) (140 tensors)
[DEBUG] stable-diffusion.cpp:432 - loading weights
[DEBUG] model.cpp:1727 - loading tensors from plcm.gguf
[INFO ] model.cpp:1897 - unknown tensor 'cond_stage_model.logit_scale | f16 | 1 [1, 1, 1, 1, 1]' in model file
[INFO ] model.cpp:1897 - unknown tensor 'cond_stage_model.text_projection | q4_0 | 2 [768, 768, 1, 1, 1]' in model file
|====> | 103/1132 - 1000.00it/ |=====> | 117/1132 - 1000.00it/ |=====> | 127/1132 - 1000.00it/ |=====> | 133/1132 - 1000.00it/ |======> | 149/1132 - 1000.00it/ |=======> | 159/1132 - 1000.00it/ |=======> | 161/1132 - 1000.00it/ |=======> | 166/1132 - 1000.00it/ |========> | 182/1132 - 1000.00it/ |========> | 187/1132 - 1000.00it/ |========> | 191/1132 - 1000.00it/ |========> | 197/1132 - 1000.00it/ |=========> | 213/1132 - 1000.00it/ |=========> | 221/1132 - 1000.00it/ |==========> | 228/1132 - 1000.00it/ |==========> | 236/1132 - 1000.00it/ |==========> | 246/1132 - 1000.00it/ |===========> | 250/1132 - 1000.00it/ |===========> | 251/1132 - 1000.00it/ |===========> | 252/1132 - 1000.00it/ |===========> | 257/1132 - 1000.00it/ |===========> | 263/1132 - 1000.00it/ |============> | 279/1132 - 1000.00it/ |============> | 292/1132 - 1000.00it/ |=============> | 301/1132 - 1000.00it/ |=============> | 315/1132 - 1000.00it/ |==============> | 334/1132 - 1000.00it/ |===================> | 444/1132 - 1000.00it/ |===================> | 451/1132 - 1000.00it/ |====================> | 453/1132 - 1000.00it/ |====================> | 455/1132 - 1000.00it/ |====================> | 460/1132 - 1000.00it/ |====================> | 463/1132 - 1000.00it/ |====================> | 465/1132 - 1000.00it/ |====================> | 467/1132 - 1000.00it/ |====================> | 469/1132 - 1000.00it/ |====================> | 470/1132 - 1000.00it/ |====================> | 472/1132 - 1000.00it/ |====================> | 474/1132 - 1000.00it/ |=====================> | 485/1132 - 1000.00it/ |=====================> | 488/1132 - 1000.00it/ |=====================> | 491/1132 - 1000.00it/ |======================> | 501/1132 - 1000.00it/ |======================> | 508/1132 - 1000.00it/ |======================> | 511/1132 - 1000.00it/ |======================> | 518/1132 - 1000.00it/ |=======================> | 522/1132 - 1000.00it/ |=======================> | 525/1132 - 1000.00it/ |=======================> | 526/1132 - 1000.00it/ |=======================> | 528/1132 - 1000.00it/ |=======================> | 538/1132 - 1000.00it/ |========================> | 553/1132 - 1000.00it/ |========================> | 556/1132 - 1000.00it/ |========================> | 563/1132 - 1000.00it/ |=========================> | 577/1132 - 1000.00it/ |=========================> | 580/1132 - 1000.00it/ |=========================> | 584/1132 - 1000.00it/ |==========================> | 593/1132 - 1000.00it/ |==========================> | 604/1132 - 1000.00it/ |==========================> | 609/1132 - 1000.00it/ |==========================> | 611/1132 - 1000.00it/ |===========================> | 622/1132 - 1000.00it/ |===========================> | 629/1132 - 1000.00it/ |============================> | 643/1132 - 1000.00it/ |============================> | 647/1132 - 1000.00it/ |============================> | 655/1132 - 1000.00it/ |=============================> | 659/1132 - 1000.00it/ |=============================> | 665/1132 - 1000.00it/ |=============================> | 669/1132 - 1000.00it/ |=============================> | 671/1132 - 1000.00it/ |==============================> | 685/1132 - 1000.00it/ |==============================> | 691/1132 - 1000.00it/ |===============================> | 705/1132 - 1000.00it/ |===============================> | 709/1132 - 1000.00it/ |===============================> | 721/1132 - 1000.00it/ |================================> | 726/1132 - 1000.00it/ |================================> | 734/1132 - 1000.00it/ |================================> | 738/1132 - 1000.00it/ |================================> | 741/1132 - 1000.00it/ |================================> | 747/1132 - 1000.00it/ |=================================> | 751/1132 - 1000.00it/ |=================================> | 761/1132 - 1000.00it/ |=================================> | 764/1132 - 1000.00it/ |=================================> | 767/1132 - 1000.00it/ |=================================> | 769/1132 - 1000.00it/ |==================================> | 782/1132 - 1000.00it/ |==================================> | 784/1132 - 1000.00it/ |==================================> | 785/1132 - 1000.00it/ |==================================> | 786/1132 - 1000.00it/ |==================================> | 787/1132 - 1000.00it/ |==================================> | 790/1132 - 1000.00it/ |==================================> | 792/1132 - 1000.00it/ |===================================> | 794/1132 - 1000.00it/ |===================================> | 796/1132 - 1000.00it/ |===================================> | 807/1132 - 1000.00it/ |===================================> | 808/1132 - 1000.00it/ |===================================> | 809/1132 - 1000.00it/ |====================================> | 820/1132 - 1000.00it/ |====================================> | 822/1132 - 1000.00it/ |====================================> | 824/1132 - 1000.00it/ |====================================> | 828/1132 - 1000.00it/ |====================================> | 830/1132 - 1000.00it/ |====================================> | 832/1132 - 1000.00it/ |====================================> | 834/1132 - 1000.00it/ |=====================================> | 841/1132 - 1000.00it/ |=====================================> | 848/1132 - 1000.00it/ |=====================================> | 852/1132 - 1000.00it/ |=====================================> | 857/1132 - 1000.00it/ |======================================> | 863/1132 - 1000.00it/ |======================================> | 865/1132 - 1000.00it/ |======================================> | 871/1132 - 1000.00it/ |=======================================> | 883/1132 - 1000.00it/ |=======================================> | 899/1132 - 1000.00it/ |=======================================> | 905/1132 - 1000.00it/ |========================================> | 907/1132 - 1000.00it/ |========================================> | 910/1132 - 1000.00it/ |=========================================> | 934/1132 - 1000.00it/ |=========================================> | 943/1132 - 1000.00it/ |=========================================> | 949/1132 - 1000.00it/ |==========================================> | 965/1132 - 1000.00it/ |==========================================> | 967/1132 - 1000.00it/ |==========================================> | 971/1132 - 1000.00it/ |===========================================> | 982/1132 - 1000.00it/ |===========================================> | 988/1132 - 1000.00it/ |===========================================> | 996/1132 - 1000.00it/ |============================================> | 997/1132 - 1000.00it/ |============================================> | 998/1132 - 1000.00it/ |============================================> | 999/1132 - 1000.00it/ |============================================> | 1000/1132 - 1000.00it |============================================> | 1001/1132 - 500.00it/ |============================================> | 1003/1132 - 500.00it/ |============================================> | 1005/1132 - 1000.00it |============================================> | 1006/1132 - 125.00it/ |============================================> | 1007/1132 - 1000.00it |============================================> | 1013/1132 - 1000.00it |============================================> | 1014/1132 - 500.00it/ |=============================================> | 1025/1132 - 1000.00it |=============================================> | 1026/1132 - 1000.00it |=============================================> | 1027/1132 - 1000.00it |=============================================> | 1031/1132 - 100.00it/ |=============================================> | 1032/1132 - 1000.00it |=============================================> | 1034/1132 - 333.33it/ |=============================================> | 1035/1132 - 500.00it/ |=============================================> | 1036/1132 - 1000.00it |=============================================> | 1038/1132 - 500.00it/ |=============================================> | 1039/1132 - 1000.00it |=============================================> | 1040/1132 - 333.33it/ |=============================================> | 1041/1132 - 1000.00it |==============================================> | 1044/1132 - 125.00it/ |==============================================> | 1046/1132 - 1000.00it |==============================================> | 1052/1132 - 500.00it/ |==============================================> | 1053/1132 - 1000.00it |==============================================> | 1062/1132 - 100.00it/ |===============================================> | 1065/1132 - 1000.00it |===============================================> | 1066/1132 - 166.67it/ |===============================================> | 1068/1132 - 142.86it/ |===============================================> | 1069/1132 - 500.00it/ |===============================================> | 1071/1132 - 500.00it/ |===============================================> | 1072/1132 - 1000.00it |===============================================> | 1073/1132 - 500.00it/ |===============================================> | 1074/1132 - 200.00it/ |===============================================> | 1075/1132 - 1000.00it |===============================================> | 1076/1132 - 500.00it/ |===============================================> | 1077/1132 - 1000.00it |===============================================> | 1078/1132 - 333.33it/ |===============================================> | 1081/1132 - 1000.00it |===============================================> | 1082/1132 - 142.86it/ |================================================> | 1087/1132 - 1000.00it |================================================> | 1092/1132 - 333.33it/ |================================================> | 1094/1132 - 1000.00it |================================================> | 1097/1132 - 500.00it/ |================================================> | 1100/1132 - 111.11it/ |================================================> | 1102/1132 - 333.33it/ |================================================> | 1103/1132 - 1000.00it |================================================> | 1106/1132 - 500.00it/ |================================================> | 1108/1132 - 500.00it/ |==================================================| 1111/1132 - 1000.00it |==================================================| 1113/1132 - 1000.00it |==================================================| 1114/1132 - 1000.00it |==================================================| 1116/1132 - 1000.00it |==================================================| 1118/1132 - 500.00it/ |==================================================| 1120/1132 - 250.00it/ |==================================================| 1121/1132 - 1000.00it |==================================================| 1122/1132 - 1000.00it |==================================================| 1125/1132 - 1000.00it |==================================================| 1130/1132 - 166.67it/ |==================================================| 1132/1132 - 31.25it/
[INFO ] stable-diffusion.cpp:531 - total params memory size = 1558.32MB (VRAM 0.00MB, RAM 1558.32MB): clip 191.00MB(RAM), unet 1272.85MB(RAM), vae 94.47MB(RAM), controlnet 0.00MB(VRAM), pmid 0.00MB(RAM)
[INFO ] stable-diffusion.cpp:535 - loading model from 'plcm.gguf' completed, taking 34.27s
[INFO ] stable-diffusion.cpp:569 - running in eps-prediction mode
[DEBUG] stable-diffusion.cpp:613 - finished loaded file
[DEBUG] stable-diffusion.cpp:1574 - txt2img 320x320
[DEBUG] stable-diffusion.cpp:1266 - prompt after extract and remove lora: "a hamster"
[INFO ] stable-diffusion.cpp:703 - Attempting to apply 0 LoRAs
[INFO ] stable-diffusion.cpp:1271 - apply_loras completed, taking 0.01s
[DEBUG] conditioner.hpp:358 - parse 'a hamster' to [['a hamster', 1], ]
[DEBUG] clip.hpp:311 - token length: 77
[DEBUG] ggml_extend.hpp:1141 - clip compute buffer size: 1.40 MB(RAM)
[DEBUG] conditioner.hpp:486 - computing condition graph completed, taking 530 ms
[DEBUG] conditioner.hpp:358 - parse '' to [['', 1], ]
[DEBUG] clip.hpp:311 - token length: 77
[DEBUG] ggml_extend.hpp:1141 - clip compute buffer size: 1.40 MB(RAM)
[DEBUG] conditioner.hpp:486 - computing condition graph completed, taking 296 ms
[INFO ] stable-diffusion.cpp:1404 - get_learned_condition completed, taking 928 ms
[INFO ] stable-diffusion.cpp:1427 - sampling using Euler A method
[INFO ] stable-diffusion.cpp:1464 - generating image: 1/1 - seed 42
[DEBUG] stable-diffusion.cpp:822 - Sample
[DEBUG] ggml_extend.hpp:1141 - unet compute buffer size: 101.29 MB(RAM)
|==================================================| 4/4 - 18.24s/it
[INFO ] stable-diffusion.cpp:1504 - sampling completed, taking 73.46s
[INFO ] stable-diffusion.cpp:1512 - generating 1 latent images completed, taking 73.61s
[INFO ] stable-diffusion.cpp:1515 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1141 - vae compute buffer size: 650.00 MB(RAM)
[DEBUG] stable-diffusion.cpp:1108 - computing vae [mode: DECODE] graph completed, taking 36.33s
[INFO ] stable-diffusion.cpp:1525 - latent 1 decoded, taking 36.33s
[INFO ] stable-diffusion.cpp:1529 - decode_first_stage completed, taking 36.33s
[INFO ] stable-diffusion.cpp:1655 - txt2img completed in 111.00s
save result PNG image to './pic.png'

=========

same command with new builds =

PS C:\Users\Amin\Desktop\AI\Koboldcpp> ./sd.exe -v -m plcm.gguf -H 320 -W 320 -o ./pic.png --steps 4 -p "a hamster"
Option:
n_threads: 4
mode: img_gen
model_path: plcm.gguf
wtype: unspecified
clip_l_path:
clip_g_path:
clip_vision_path:
t5xxl_path:
diffusion_model_path:
high_noise_diffusion_model_path:
vae_path:
taesd_path:
esrgan_path:
control_net_path:
embedding_dir:
stacked_id_embed_dir:
input_id_images_path:
style ratio: 20.00
normalize input image: false
output_path: ./pic.png
init_image_path:
end_image_path:
mask_image_path:
control_image_path:
ref_images_paths:
increase_ref_index: false
offload_params_to_cpu: false
clip_on_cpu: false
control_net_cpu: false
vae_on_cpu: false
diffusion flash attention: false
diffusion Conv2d direct: false
vae_conv_direct: false
control_strength: 0.90
prompt: a hamster
negative_prompt:
clip_skip: -1
width: 320
height: 320
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: default, sample_method: euler_a, sample_steps: 4, eta: 0.00)
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: default, sample_method: euler_a, sample_steps: -1, eta: 0.00)
moe_boundary: 0.875
flow_shift: inf
strength(img2img): 0.75
rng: cuda
seed: 42
batch_count: 1
vae_tiling: false
upscale_repeats: 1
chroma_use_dit_mask: true
chroma_use_t5_mask: false
chroma_t5_mask_pad: 1
video_frames: 1
fps: 16
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] stable-diffusion.cpp:174 - Using CPU backend
PS C:\Users\Amin\Desktop\AI\Koboldcpp>

[wbruna]

So this release is already crashing in that way: https://github.com/leejet/stable-diffusion.cpp/releases/tag/master-0d8b39f ? Please check it carefully, because I can't see how those code changes could cause such a behavior (unless master-0d8b39f is working, but master-ecf5db9 isn't?).

I could also try that exact model, since it looks like the crash happens right before it is loaded. I see it's a q4_0 quant; did you convert it yourself? Could you point me to where I could download it, or the corresponding .safetensors file?

[Amin456789]

No please check the link instead of the number

Newer builds are not working

It doesn't matter which model. I tried wan and the other stuff too. It's not the model I use. It just crash with any model even chroma. The gguf I have works great with local diffusion on Android and koboldpp too

[Amin456789]

Edited my -v post to know what I mean

[Amin456789]

The build that has added opencl is all started the crashes. Maybe that is the culprit

[wbruna]

No please check the link instead of the number

It is master-b1fc16b , isn't it? That's why I asked about the release master-0d8b39f : that's the first release after master-b1fc16b (the order on the release page is wrong; I'm following the code history).

The build that has added opencl is all started the crashes. Maybe that is the culprit

master-d42fd59 , which added OpenCL support, is one after master-0d8b39f . It is more likely to cause a crash, because the ggml library was updated too; but please, just confirm: is master-0d8b39f crashing, or not?

Following the code order, we have:

• b1fc16b - resetting clip_skip - definitely works
• 539b5b9 - musa docker build (no binary release)
• 0d8b39f - avoid crash on sdxl loras - ???
• d42fd59 - OpenCL - crashes
• 23de7fc - build warnings on Linux - crashes
• ea46fd6 - zero-initialize output of tiling (no binary release)
• ecf5db9 - fix windows build - crashes

It doesn't matter which model. I tried wan and the other stuff too. It's not the model I use. It just crash with any model even chroma. The gguf I have works great with local diffusion on Android and koboldpp too

Just a wild guess: could you check if the avx binaries crash in the same way?

[Amin456789]

yes master-0d8b39f works great with no problem, after that everything else just crashes, so i think opencl is doing something, can we disable it or have an option force cpu maybe for test with latest version please?

yes i tested every build from all avx and no avx too, its all crashes

[wbruna]

I don't think OpenCL support itself could cause this, because it's a build-time option, disabled by default (and it'd appear as "Using OpenCL backend" if it were enabled by mistake: d42fd59 ).

But the backend initialization is done by ggml, and that same revision updated the ggml submodule, which had a ton of pending changes: ggml-org/ggml@ff90529...9e4bee1

So one possibility would be a change on ggml. Except... you mentioned Koboldcpp works. Does recent Koboldcpp releases work for you? Because it also uses the same ggml library, updating it far more frequently actually.

[wbruna]

By the way, any chance you are running sd.exe with an unmatched stable-diffusion.dll (from another release, for instance)? Even that shouldn't cause problems between these two releases, but...