An error occurs when launching the benchmark.

[shiwk]

RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same

The issue arises when running the benchmark code for the Vision Transformer (VIT). I believe it occurs because the input tensor is located on the CPU, while the model is on the GPU.

It will work with the addition of a single line of code:
model = self.model.to(self.device)

[IsaevIlya]

Hello @shiwk,
Thank you for letting us know about the issue. We will have a look on it. Meanwhile, could you share the instance type on which you faced that issue?

[shiwk]

Hello @shiwk, Thank you for letting us know about the issue. We will have a look on it. Meanwhile, could you share the instance type on which you faced that issue?

Hi @IsaevIlya , I run it on my own server with a GPU, not on any service of AWS

[IsaevIlya]

@shiwk, Thank you for bringing up this issue. I've successfully replicated it, and the resolution should be addressed by the PR #172