[Admin maintenance] Support new ZeroGPU hardware

Thank you so much for having shared this Space with the community on this demo. We have upgraded the ZeroGPU infra-structure to run on modern blackwell architecture.
For that, we need to upgrade your demo to support that. This PR fixes your demo to work with the new architecture. As this is something we broke on our end, we may merge this PR autonomously. If this breaks unexpectedly or brings unintended consequences, feel free to revert, modify or otherwise. Any issues you can email apolinario@huggingface.co

README.md

@@ -4,7 +4,7 @@ emoji: 🐨
 colorFrom: indigo
 colorTo: pink
 sdk: gradio
-sdk_version: 4.37.1
+sdk_version: 6.20.0
 app_file: app.py
 pinned: false
 license: apache-2.0

app.py

@@ -1,14 +1,31 @@
+import spaces
 import os
+# The 2.0 rewrite of DiffSynth-Studio restructured the package into diffsynth/core/...
+# and dropped ModelManager/SVDVideoPipeline (and requirements.txt), breaking the runtime
+# `git clone` + `pip install -r DiffSynth-Studio/requirements.txt` this Space used to do.
+# Pin the clone to the last pre-2.0 commit. All of DiffSynth's own deps (torch, transformers,
+# huggingface_hub, einops, etc.) are now pinned directly in this Space's requirements.txt and
+# installed at BUILD time, so at runtime we only need to fetch and copy the plain-Python
+# `diffsynth` package itself (no pip install here: `import spaces` above already pulled
+# huggingface_hub into memory, so a runtime `pip install --upgrade` would update the files on
+# disk but not the module object already cached in this process's sys.modules).
 os.system("git clone https://github.com/modelscope/DiffSynth-Studio.git")
+os.system("cd DiffSynth-Studio && git checkout afd101f3452c9ecae0c87b79adfa2e22d65ffdc3")
 os.system("cp -r DiffSynth-Studio/diffsynth ./")
-os.system("pip install -r DiffSynth-Studio/requirements.txt")
+# gradio 6 needs huggingface_hub>=1.2, but this pre-2.0 DiffSynth checkout needs a
+# huggingface_hub new enough to satisfy modern transformers (CachedRepoTreeNotFoundError) --
+# together that only resolves with a 5.x transformers, which dropped the long-deprecated
+# re-export `transformers.modeling_utils.PretrainedConfig` that diffsynth's
+# stepvideo_text_encoder.py still imports. Shim it back before diffsynth is imported.
+import transformers.modeling_utils as _modeling_utils
+from transformers.configuration_utils import PretrainedConfig as _PretrainedConfig
+if not hasattr(_modeling_utils, "PretrainedConfig"):
+    _modeling_utils.PretrainedConfig = _PretrainedConfig
 from diffsynth import save_video, ModelManager, SVDVideoPipeline
-from diffsynth import ModelManager
-import torch, os, random, time
+import torch, random, time
 import gradio as gr
 import numpy as np
 from PIL import Image
-import spaces
 
 
 def get_i2v_pipeline():
@@ -99,7 +116,7 @@ To use this model, please refer to [DiffSynth](https://github.com/modelscope/Dif
         inputs=image,
         outputs=[video, seed],
         fn=process_examples,
-        cache_examples="lazy",
+        cache_examples=False,
     )
 
 if __name__ == "__main__":

requirements.txt

@@ -0,0 +1,16 @@
+torch==2.11.0
+torchvision==0.26.0
+huggingface_hub==1.23.0
+transformers
+imageio
+imageio[ffmpeg]
+safetensors
+einops
+sentencepiece
+protobuf
+modelscope
+ftfy
+pynvml
+pandas
+accelerate
+peft