[Admin maintenance] Support new ZeroGPU hardware

README.md

@@ -4,7 +4,7 @@ emoji: 🦀
 colorFrom: indigo
 colorTo: red
 sdk: gradio
-sdk_version: 4.0.0
+sdk_version: 5.49.1
 app_file: app.py
 pinned: false
 license: mit

opensora/serve/gradio_web_server.py

@@ -5,6 +5,39 @@ import sys
 import os
 import random
 
+import spaces
+
+# diffusers 0.24 imports `HfFolder` and `cached_download` from huggingface_hub
+# at module load. Both were removed in hub 0.26 / 1.x. Stub them so the import
+# succeeds; the demo path doesn't actually exercise these functions.
+import huggingface_hub as _hf_hub
+if not hasattr(_hf_hub, "cached_download"):
+    def _cached_download_removed(*a, **k):
+        raise NotImplementedError(
+            "huggingface_hub.cached_download was removed; bump diffusers or pin hub<0.26."
+        )
+    _hf_hub.cached_download = _cached_download_removed
+if not hasattr(_hf_hub, "HfFolder"):
+    class _HfFolderStub:
+        @staticmethod
+        def get_token():
+            import os
+            return os.environ.get("HF_TOKEN")
+    _hf_hub.HfFolder = _HfFolderStub
+
+# Force xformers to use the Cutlass kernel — the auto-picked FA3 path is
+# Hopper-only and crashes on Blackwell (sm_120) with "CUDA error ... invalid argument".
+try:
+    import xformers.ops as _xops
+    _orig_mea = _xops.memory_efficient_attention
+    _cutlass_ops = (_xops.fmha.cutlass.FwOp, _xops.fmha.cutlass.BwOp)
+    def _mea(*a, **kw):
+        kw.setdefault("op", _cutlass_ops)
+        return _orig_mea(*a, **kw)
+    _xops.memory_efficient_attention = _mea
+except Exception:
+    pass
+
 import imageio
 import torch
 from diffusers import PNDMScheduler
@@ -24,8 +57,6 @@ from opensora.models.diffusion.latte.modeling_latte import LatteT2V
 from opensora.sample.pipeline_videogen import VideoGenPipeline
 from opensora.serve.gradio_utils import block_css, title_markdown, randomize_seed_fn, set_env, examples, DESCRIPTION
 
-import spaces
-
 @spaces.GPU(duration=300)
 def generate_img(prompt, sample_steps, scale, seed=0, randomize_seed=False, force_images=False):
     seed = int(randomize_seed_fn(seed, randomize_seed))

requirements.txt

@@ -1,12 +1,12 @@
-torch==2.1.2
-torchvision==0.16.2
+torch==2.8.0
+torchvision==0.23.0
 transformers==4.39.1
 accelerate==0.28.0
 albumentations==1.4.0
-av==11.0.0
+av
 decord==0.6.0
 einops==0.7.0
-fastapi==0.110.0
+fastapi
 gdown==5.1.0
 h5py==3.10.0
 idna==3.6
@@ -33,7 +33,7 @@ torchdiffeq==0.2.3
 torchmetrics==1.3.2
 tqdm==4.66.2
 urllib3==2.2.1
-uvicorn==0.27.1
+uvicorn
 diffusers==0.24.0
 scikit-video==1.1.11
 imageio-ffmpeg==0.4.9
@@ -43,5 +43,3 @@ ftfy==6.1.3
 moviepy==1.0.3
 wandb==0.16.3
 tensorboard==2.14.0
-pydantic==2.6.4
-deepspeed==0.12.6