Float8Tensor dispatch error: aten.as_strided not implemented in DenoiseBlock

[livepeer-robot]

Summary

The DenoiseBlock fails when using Float8 quantization (via torchao) because the aten.as_strided operator is not implemented for Float8Tensor.

Error Details

ERROR log from fal.ai backend (20 Feb 2026 11:23 UTC):

Error in block: (denoise, DenoiseBlock)
Error details: Float8Tensor dispatch: attempting to run unimplemented operator/function: func=<OpOverload(op='aten.as_strided', overload='default')>, types=(<class 'torchao.quantization.Float8Tensor'>,), arg_types=(<class 'torchao.quantization.Float8Tensor'>, <class 'list'>, <class 'list'>, <class 'int'>), kwarg_types={}

Analysis

• torchao.quantization.Float8Tensor does not support the as_strided PyTorch operation
• This is called during the denoising process, likely during attention or tensor reshaping
• The issue occurs with Float8 quantized models

Possible Solutions

1. Workaround: Disable Float8 quantization for affected operations
2. Upstream fix: Contribute as_strided support to torchao's Float8Tensor
3. Guard clause: Detect Float8Tensor and use alternative tensor operations that are supported

Environment

• Pipeline: LongLive (DenoiseBlock)
• Quantization: Float8 via torchao
• Timestamp: 2026-02-20 11:23:33 UTC

References

• torchao Float8 implementation: https://github.com/pytorch/ao
• This may require upstream changes to torchao's Float8Tensor.__torch_dispatch__