Skip to content

sd: sync with master-721-8caa3f9#2284

Merged
LostRuins merged 6 commits into
LostRuins:concedo_experimentalfrom
wbruna:kcpp_sd_update_202606_6
Jun 25, 2026
Merged

sd: sync with master-721-8caa3f9#2284
LostRuins merged 6 commits into
LostRuins:concedo_experimentalfrom
wbruna:kcpp_sd_update_202606_6

Conversation

@wbruna

@wbruna wbruna commented Jun 21, 2026

Copy link
Copy Markdown
7837596-a cute kobold wearing a shirt with boogu written on it

master-714-b12098f:

  • Boogu-Image support
  • fixes for Anima and Ernie with Vulkan + Flash Attention
  • a CLIP normalization fix
  • dropped the assertion failure for SD_TYPE_COUNT != GGML_TYPE_COUNT, which shouldn't be needed anymore
  • enabled the mmproj for LongCat-Image Edit (although I haven't had much luck with it)

@wbruna

wbruna commented Jun 21, 2026

Copy link
Copy Markdown
Author

I'm hitting an assertion failure with Boogu-Image Edit:

ggml/src/ggml.c:2612: GGML_ASSERT(a->ne[d] == b->ne[d]) failed

so I've disabled its edit mode for now.

Edit: I can't reproduce the crash with stable-diffusion.cpp.

@wbruna wbruna force-pushed the kcpp_sd_update_202606_6 branch from 4a7f9e3 to 31e05bf Compare June 22, 2026 02:21
@wbruna

wbruna commented Jun 22, 2026

Copy link
Copy Markdown
Author

Oddly, after removing the mmproj (I guess Qwen3-VL has it built-in), and reducing cfg to 1, I've got the edit mode working:

8952225-change the text boogu in the shirt to edit mode

@LostRuins

Copy link
Copy Markdown
Owner

does it need any tensor loading hacks?

@wbruna

wbruna commented Jun 22, 2026

Copy link
Copy Markdown
Author

Apart from the llm and mmproj redirects, I didn't need any. But I didn't try a downloaded quant; for the images above I've used q8_0 quants I've converted myself (with a simple --type q8_0). Working both on Vulkan and ROCm.

@wbruna wbruna force-pushed the kcpp_sd_update_202606_6 branch from 31e05bf to e7c5d6a Compare June 22, 2026 22:02
@wbruna

wbruna commented Jun 22, 2026

Copy link
Copy Markdown
Author

master-719-f440ad9:

  • mmap fix
  • guidance_schedule parameter (useful for Ideogram once supported)
  • eager_load parameter to preload weights (but doesn't seem to work without streaming)

@wbruna wbruna changed the title sd: sync with master-714-b12098f sd: sync with master-719-f440ad9 Jun 23, 2026
@wbruna wbruna force-pushed the kcpp_sd_update_202606_6 branch from e7c5d6a to 5a45265 Compare June 24, 2026 23:11
@wbruna wbruna changed the title sd: sync with master-719-f440ad9 sd: sync with master-721-8caa3f9 Jun 24, 2026
@wbruna

wbruna commented Jun 24, 2026

Copy link
Copy Markdown
Author
1-a small orange kobold, pointing to a blackboard where it is written Krea 2_tae

master-721-8caa3f9:

  • Krea 2 support
  • logit-normal scheduler

@LostRuins LostRuins merged commit afbd83b into LostRuins:concedo_experimental Jun 25, 2026
@wbruna

wbruna commented Jun 25, 2026

Copy link
Copy Markdown
Author

@LostRuins , I'm getting some failures in image to image mode with TAE. I think that depending on the VAE type and resolution, TAE also needs tiling to avoid OOMs. Maybe we could switch it to follow sdtiledvae with an extra multiplier (say, 2x the area)?

@LostRuins

Copy link
Copy Markdown
Owner

@wbruna i thought tiling was not supported on TAE? If it's supported we can probably pass it through for larger images yes

@wbruna

wbruna commented Jun 26, 2026

Copy link
Copy Markdown
Author

@wbruna i thought tiling was not supported on TAE? If it's supported we can probably pass it through for larger images yes

@LostRuins , stduhpf unified those paths some time ago.

But it's not the resolution: ( Krea2 or Anima ) + i2i + sdvaeauto is crashing even with tiny images. Also tried replacing taesd_w21.embd with the original .safetensors file.

Edit: upstream issue.

Edit 2: I think I've found the bug; I'll send a patch later.

@LostRuins

Copy link
Copy Markdown
Owner

cool thanks

@wbruna

wbruna commented Jun 26, 2026

Copy link
Copy Markdown
Author

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants