AutoProcessor import fails with NVIDIA torch 2.7 missing torch.float8_e8m0fnu

[doctorpangloss]

System Info

• transformers: 5.10.1
• torch: 2.7.0a0+7c8ec84dab.nv25.03
• Python: 3.12
• CUDA runner: CUDA 12.8.1.012 / NVIDIA driver 580.126.09

Transformers metadata still lists torch>=2.4 for the torch/testing extras.

Reproduction

In an environment with transformers==5.10.1 and NVIDIA torch 2.7.0a0+7c8ec84dab.nv25.03:

import torch
print(torch.__version__)
print(hasattr(torch, "float8_e8m0fnu"))

from transformers import AutoProcessor

Result:

2.7.0a0+7c8ec84dab.nv25.03
False
AttributeError: module 'torch' has no attribute 'float8_e8m0fnu'

Full import path from the traceback:

from transformers import AutoProcessor
transformers/models/auto/processing_auto.py
transformers/image_processing_utils.py
transformers/processing_utils.py
transformers/modeling_utils.py
transformers/integrations/finegrained_fp8.py: _UE8M0_SF_DTYPE = torch.float8_e8m0fnu

I also checked the public PyTorch CPU torch==2.7.0 wheel and it has this dtype, so the failing case is the NVIDIA torch 2.7 build used by the CUDA runner.

Expected behavior

Importing AutoProcessor should not fail just because this optional finegrained FP8 dtype is missing. Either guard this dtype access/import until the FP8 integration is used, or require a torch version/build that provides torch.float8_e8m0fnu.

[Rocketknight1]

cc @ydshieh @tarekziade maybe, since this seems to be an issue with a specific test build? Not sure exactly what's going on

[Rocketknight1]

Actually nvm there's already reviews on the fix at #46393! Trusting them