We are hitting a reproducible crash in shot_switching_extension while integrating SkyReels V3 into a production Modal runtime.
What works in the same environment:
reference_to_video
talking_avatar
single_shot_extension
What fails:
Environment
- Repo commit:
28c771e8456341be6a213e3d1133ed1fd19bf75d
- GPU: H100
- Runtime: Modal container
- Model:
Skywork/SkyReels-V3-Video-Extension
- We invoke
generate_video.py with --task_type shot_switching_extension ... --offload
Observed behavior
The request is accepted, the input video is staged correctly, the model loads, and execution enters the shot switching pipeline. It then crashes inside the transformer path.
Relevant log excerpt
2026-03-13 08:10:00 - skyreels_v3 - INFO - [generate_video.py:274 - <module>] - input params: Namespace(task_type='shot_switching_extension', model_id='/cache/hub/models--Skywork--SkyReels-V3-Video-Extension/snapshots/68def6119d5731ab5f11faeda18a4a6bf65ff52e', duration=5, prompt='<redacted>.', resolution='720P', seed=42, use_usp=False, offload=True, low_vram=False, input_video='/inputs/holoscript/7160c867-dfe8-439b-b6b8-92a609649f0f-seg-1-retry-0/input_video.mp4', ...)
...
Traceback (most recent call last):
File "/root/SkyReels-V3/generate_video.py", line 286, in <module>
video_out = pipe.extend_video(args.input_video, args.prompt, args.duration, args.seed, resolution=args.resolution)
File "/root/SkyReels-V3/skyreels_v3/pipelines/shot_switching_extension_pipeline.py", line 114, in extend_video
video_frames = self.__call__(
File "/usr/local/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 120, in decorate_context
return func(*args, **kwargs)
File "/root/SkyReels-V3/skyreels_v3/pipelines/shot_switching_extension_pipeline.py", line 217, in __call__
noise_pred = self.transformer(
File "/usr/local/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
return forward_call(*args, **kwargs)
File "/root/SkyReels-V3/skyreels_v3/modules/transformer.py", line 738, in forward
if not block_offload and self.freqs.device != x.device and isinstance(x, torch.Tensor):
AttributeError: 'list' object has no attribute 'device'
Using the same environment and the same model family, single_shot_extension succeeds on our source videos. This suggests the failure is specific to the shot_switching_extension code path rather than a general deployment/configuration problem.
We are hitting a reproducible crash in
shot_switching_extensionwhile integrating SkyReels V3 into a production Modal runtime.What works in the same environment:
reference_to_videotalking_avatarsingle_shot_extensionWhat fails:
shot_switching_extensionEnvironment
28c771e8456341be6a213e3d1133ed1fd19bf75dSkywork/SkyReels-V3-Video-Extensiongenerate_video.pywith--task_type shot_switching_extension ... --offloadObserved behavior
The request is accepted, the input video is staged correctly, the model loads, and execution enters the shot switching pipeline. It then crashes inside the transformer path.
Relevant log excerpt
Using the same environment and the same model family,
single_shot_extensionsucceeds on our source videos. This suggests the failure is specific to theshot_switching_extensioncode path rather than a general deployment/configuration problem.