Same model gets slightly different results on different platforms

1

The model gives silightly different results on different platforms. And the model is trained on GPU and save as .cuda() format.
Is there a way to make the result to be precisely the same?
with torch 1.13.1+cu116, on Tesla A100
torch 1.13.1+cu116, on GTX1060
and torch 2.1.2 on MacOS mps

The test script goes like this

    def test_model_consistency(self):
        if(torch.cuda.is_available()):
            device = torch.device("cuda")
        elif(torch.backends.mps.is_available()):
            device = torch.device("mps")
        else:
            device = torch.device("cpu")
        model_path = "best.pth"
        if(torch.cuda.is_available()):
            model = torch.load(model_path).to(device)
        else:
            model = torch.load(model_path, map_location=device).to(device)
        output = model(torch.zeros(1,128,6).to(device))
        print(output)

And I get
image-20231229204928278

image-20231229205035341
image-202312292050353411108×295 21.7 KB

232f7e131c759d9a51ab89f0e3233d88
232f7e131c759d9a51ab89f0e3233d88888×216 54.1 KB

There are slightly differences for the output, but I want them to be exactly the same. Is there any way to do that?

2

No, since different hardware and their corresponding libraries might use different algorithms causing the expected, small numerical mismatches caused due to the limited floating point precision.