Testing : V10.61 : some issues In Kandinsky 5 (Fix in pipeline_hunyuan_video.py)`

[utilewebsites]

I am reporting an issue regarding Kandinsky 5 V10.61. I have set up the venv according to the latest requirements.

I encounter the following error context:

"a 'config.json' that describes the model is required in the directory of the model or inside the safetensor file"

The Fix:
Since December, I have to manually patch /models/hyvideo/diffusion/pipelines/pipeline_hunyuan_video.py to prevent a crash. It seems ref_latents can be None, causing the index call to fail.

Location: Line 1661

Original Code:

ref_latents=ref_latents[j].unsqueeze(0),

My Fix (working since Dec):

ref_latents=None if ref_latents is None else ref_latents[j].unsqueeze(0),

Could this check be included in the official code? I am curious if I am the only one needing this patch for the pipeline to work.

Thanks!

---

System : Ubuntu 24.04 (3090)

[NigYoA]

I tried to follow with this solution, but it does not work for me. shouldn’t this folder (or similar setting folder/file) be inside Kandinsky instead of hyvideo?

File "G:\pinokio\api\Wan2GP\app\wgp.py", line 6900, in generate_video_error_handler
generate_video(task, send_cmd, plugin_data=plugin_data, **params)
File "G:\pinokio\api\Wan2GP\app\wgp.py", line 5607, in generate_video
wan_model, offloadobj = load_models(
File "G:\pinokio\api\Wan2GP\app\wgp.py", line 3491, in load_models
wan_model, pipe = model_type_handler.load_model(
File "G:\pinokio\api\Wan2GP\app\models\kandinsky5\kandinsky_handler.py", line 282, in load_model
kandinsky = model_factory(
File "G:\pinokio\api\Wan2GP\app\models\kandinsky5\kandinsky_main.py", line 152, in init
text_embedder = get_text_embedder(
File "G:\pinokio\api\Wan2GP\app\models\kandinsky5\kandinsky\models\text_embedders.py", line 284, in get_text_embedder
return Kandinsky5TextEmbedder(conf, device, quantized_qwen, text_token_padding)
File "G:\pinokio\api\Wan2GP\app\models\kandinsky5\kandinsky\models\text_embedders.py", line 228, in init
self.clip_embedder = ClipTextEmbedder(conf.clip, device)
File "G:\pinokio\api\Wan2GP\app\models\kandinsky5\kandinsky\models\text_embedders.py", line 19, in init
self.model= offload.fast_load_transformers_model( os.path.join(conf.checkpoint_path,"model.safetensors"), modelClass=CLIPTextModel, ignore_unused_weights= True, forc
edConfigPath = os.path.join(conf.checkpoint_path, "text_config.json"))
File "G:\pinokio\api\Wan2GP\app\env\lib\site-packages\mmgp\offload.py", line 1586, in fast_load_transformers_model
raise Exception("a 'config.json' that describes the model is required in the directory of the model or inside the safetensor file")
Exception: a 'config.json' that describes the model is required in the directory of the model or inside the safetensor file

[utilewebsites]

Hi @NigYoA,

My fix is specifically for the Hunyuan pipeline, so it does not apply to Kandinsky.

Looking at your error log, the issue seems to be in text_embedders.py, which indicates that your Kandinsky model folder is incomplete. You are likely missing text_config.json (or the folder structure is incorrect) in G:\pinokio\api\Wan2GP\app\models\kandinsky5.