Error on multi-GPU, likely caused by PyTorch 2.7

[BenjaminBossan]

Hi, the AWQ multi-GPU unit test in PEFT started failing recently with the error:

ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)

The full test, which uses Trainer for DP, can be found here. I could reproduce the error locally using PyTorch 2.7. When switching to an env with PyTorch 2.6, the test passes, so I think this is the most likely culprit.