fix sd35 with batch

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -3,8 +3,9 @@
 ## Update for 2025-03-17
 
 ### TODO
-  - Gemma3 requires `git+https://github.com/huggingface/transformers@v4.49.0-Gemma-3`
+  - Gemma3 requires `transformers==git+https://github.com/huggingface/transformers@v4.49.0-Gemma-3`
   - Remote VAE encode for SD15 and Flux.1: <https://github.com/huggingface/diffusers/issues/11069>
+  - HunyuanVideo-I2V: <https://github.com/huggingface/diffusers/issues/11118>
 
 ### Highlights for 2025-03-17
 
@@ -50,6 +51,7 @@ Support for [CogView 4](https://huggingface.co/THUDM/CogView4-6B), new CLiP mode
   - add quantization support to **CogView-3Plus**
   - update `diffusers` and other requirements   
   - rename vae, unet and text-encoder settings *None* to *Default* to avoid confusion  
+  - **CLI**: add `cli/api-grid.py` which can generate grids using params-from-file for x/y axis  
 - **IPEX**
   - add `--upgrade` to torch_command when using `--use-nightly`  
   - add xpu to profiler  

modules/processing_vae.py

@@ -293,9 +293,6 @@ def vae_decode(latents, model, output_type='np', vae_type='Full', width=None, he
         latents = model._unpack_latents(latents, height, width, model.vae_scale_factor) # pylint: disable=protected-access
     if len(latents.shape) == 3: # lost a batch dim in hires
         latents = latents.unsqueeze(0)
-    if latents.shape[0] == 4 and latents.shape[1] != 4: # likely animatediff latent
-        latents = latents.permute(1, 0, 2, 3)
-
     if latents.shape[-1] <= 4: # not a latent, likely an image
         decoded = latents.float().cpu().numpy()
     elif vae_type == 'Full' and hasattr(model, "vae"):

modules/video_models/hunyuan.py

@@ -63,6 +63,12 @@ def hijack_encode_prompt(*args, **kwargs):
     return res
 
 
+def get_quant(args):
+    if args is not None and "quantization_config" in args:
+        return args['quantization_config'].__class__.__name__
+    return None
+
+
 def load(selected):
     if selected is None:
         return
@@ -75,7 +81,7 @@ def load(selected):
     quant_args = model_quant.create_config(module='Model')
     cls = diffusers.HunyuanVideoTransformer3DModel
     try:
-        debug(f'Video load: module=transofrmer repo="{selected.dit}" subfolder="{selected.subfolder}" cls={cls.__name__} quant={quant_args is not None}')
+        debug(f'Video load: module=transofrmer repo="{selected.dit}" subfolder="{selected.subfolder}" cls={cls.__name__} quant={get_quant(quant_args)}')
         transformer = cls.from_pretrained(
             pretrained_model_name_or_path=selected.dit,
             subfolder=selected.subfolder,
@@ -92,7 +98,7 @@ def load(selected):
     else:
         cls = transformers.LlamaModel
     try:
-        debug(f'Video load: module=te repo="{selected.repo}" cls={cls.__name__} quant={quant_args is not None}')
+        debug(f'Video load: module=te repo="{selected.repo}" cls={cls.__name__} quant={get_quant(quant_args)}')
         text_encoder = cls.from_pretrained(
             pretrained_model_name_or_path=selected.repo,
             subfolder="text_encoder",
@@ -107,7 +113,7 @@ def load(selected):
 
     cls = transformers.CLIPTextModel
     try:
-        debug(f'Video load: module=clip repo="{selected.repo}" cls={cls.__name__} quant=False')
+        debug(f'Video load: module=clip repo="{selected.repo}" cls={cls.__name__} quant=None')
         text_encoder_2 = transformers.CLIPTextModel.from_pretrained(
             pretrained_model_name_or_path=selected.repo,
             subfolder="text_encoder_2",
@@ -119,7 +125,7 @@ def load(selected):
 
     cls = diffusers.AutoencoderKLHunyuanVideo
     try:
-        debug(f'Video load: module=vae repo="{selected.repo}" cls={cls.__name__} quant=False')
+        debug(f'Video load: module=vae repo="{selected.repo}" cls={cls.__name__} quant=None')
         vae = diffusers.AutoencoderKLHunyuanVideo.from_pretrained(
             pretrained_model_name_or_path=selected.repo,
             subfolder="vae",
@@ -136,7 +142,7 @@ def load(selected):
     else:
         cls = diffusers.HunyuanVideoPipeline
     try:
-        debug(f'Video load: module=pipe repo="{selected.repo}" cls={cls.__name__} quant=False')
+        debug(f'Video load: module=pipe repo="{selected.repo}" cls={cls.__name__} quant=None')
         shared.sd_model = cls.from_pretrained(
             pretrained_model_name_or_path=selected.repo,
             transformer=transformer,
@@ -205,6 +211,7 @@ def generate(*args, **kwargs):
             shared.log.error('Video: init image not set')
             return [], None, '', '', 'Error: init image not set'
         p.task_args['image'] = init_image
+        # from PIL import Image
         # p.task_args['image'] = init_image.resize((336, 336), Image.Resampling.LANCZOS)
 
     shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model)