nunchaku

nunchaku copied to clipboard

Can svdquant comfyui be used with flux pulid?

9

model_type FLUX [2024-12-25 14:49:43.765] [info] Loading partial weights from /home/ubuntu/.cache/huggingface/hub/models--mit-han-lab--svdquant-models/snapshots/0684a6e2693230e1aa0c54827a90e64bf082c95d/svdq-flux.1-dev-lora-realism.safetensors [2024-12-25 14:49:43.765] [warning] Unable to pin memory: operation not supported [2024-12-25 14:49:43.765] [info] Try MAP_PRIVATE [2024-12-25 14:49:43.861] [info] Done. [2024-12-25...

pyoliu

D:\ollmamodel\nunchaku\nunchaku\third_party\spdlog\include\spdlog\fmt\bundled\format.h(3273): warning #27-D: character value is out of range return U"\x9999999a\x828f5c29\x80418938\x80068db9\x8000a7c6\x800010c7" ^ D:\ollmamodel\nunchaku\nunchaku\third_party\spdlog\include\spdlog\fmt\bundled\format.h(3273): warning #27-D: character value is out of range return U"\x9999999a\x828f5c29\x80418938\x80068db9\x8000a7c6\x800010c7" ^ D:\ollmamodel\nunchaku\nunchaku\third_party\spdlog\include\spdlog\fmt\bundled\format.h(3274): warning #27-D: character value is...

swh2026

AssertionError: assert image_rotary_emb.shape[2] == batch_size * (txt_tokens + img_tokens)

5

I get the following error when execute inpainting with batch size 1, using flux-fill workflow, ``` Traceback (most recent call last): File "/root/ComfyUI/execution.py", line 327, in execute output_data, output_ui, has_subgraph...

O-O1024

Collecting nunchaku==0.2.0+torch2.6 Downloading https://huggingface.co/mit-han-lab/nunchaku/resolve/main/nunchaku-0.2.0+torch2.6-cp310-cp310-win_amd64.whl (125.3 MB) ---------------------------------------- 125.3/125.3 MB 42.6 MB/s eta 0:00:00 Requirement already satisfied: pip in d:\comfyui\comfyui-aki-v1.3\python\lib\site-packages (24.2) ERROR: Could not find a version that satisfies the requirement...

gdx09090909

[Bug] t5 encoder problem

8

i use it in flux comfyui workflow, the textencoder node seems needing three files,clip-L t5 and svdq-t5. The memory usage is huge,if we can only use clip-L.safetensor and svdq-t5.safetensors(3G file...

zwukong

Flux FP8 Lora from sd-scripts not working

2

Trained lora using https://github.com/kohya-ss/sd-scripts and converting it to fp8 using same scripts works fine with the flux-dev-fp8 model. With nunchacku fp4 it fails with the following error. Could you add...

dimonnwc3

Anyway to disable set_attention_impl and apply_cache_on_pipe

4

Hello, Many congratulations on the release. Some really good features, thank you. I am integration performance improvements in custom Gradio UI. I basically reuse the pipe for subsequent generations. Now...

nitinmukesh

[Feature] 请问有计划支持分区lora绘制吗？目前使用会报错。

2

首先感谢团队做出的卓越贡献，controlnet应用速度大幅度提升了。但是目前貌似teacache等加速技术就失效了，使用时没看到明显的效果。 另外分区绘制的时候如图，单纯只是接入Nunchaku FLUX DiT Loader这个节点作为开始都会报错，这个分区绘制速度特别慢，特别需要加速: The size of tensor a (266) must match the size of tensor b (256) at non-singleton dimension 1 

tensound

The same set of prompt words, the results of the generation are inconsistent

3

GPU5070ti,The other parameters of the same set of prompt words are the same, but there are subtle differences in the results. This will not happen if the model with fp8...

CGF985

how to convert other flux checkpoint?

1

how to convert other flux checkpoint? for example: https://huggingface.co/shuttleai/shuttle-jaguar/tree/main https://civitai.com/models/1090420/edge-of-reality

xueqing0622