app.py

import gradio as gr
import spaces
import torch

torch.set_default_device('cuda')


class LinearModel(torch.nn.Module):
    def __init__(self):
        super(LinearModel, self).__init__()
        self.linear = torch.nn.Linear(1, 1, device='cuda')

    def forward(self, x):
        out = self.linear(x)
        print('weight device: ' + str(self.linear.weight.device))
        return out


def process():
    print('cuda avaliable: ' + str(torch.cuda.is_available()))
    print('cuda device count: ' + str(torch.cuda.device_count()))

    model = LinearModel().cuda()
    x = torch.ones(1, device='cuda')
    y = model(x)

    print(model)
    print(x)
    print(y)


@spaces.GPU
def greet(n):
    process()
    return f"Hello {n} Tensor"


def func(n):
    # step1, on cpu
    process()

    # step2, on gpu
    res = greet(n)

    # step3, on cpu
    process()

    return res


gr.Interface(fn=func, inputs=gr.Number(), outputs=gr.Text()).launch()