Test ZeroGPU

app.py

@@ -2,13 +2,29 @@ import gradio as gr
 import spaces
 import torch
 
-zero = torch.Tensor([0]).cuda()
-print(zero.device) # <-- 'cpu' 🤔
+# Outside of GPU context → no GPU available
+zero = torch.tensor([0.0])  # stays on CPU at import time
+print("startup device:", zero.device)  # will log: cpu
 
-@spaces.GPU
-def greet(n):
-    print(zero.device) # <-- 'cuda:0' 🤗
-    return f"Hello {zero + n} Tensor"
+@spaces.GPU  # GPU is allocated ONLY while this function runs
+def greet(n: float):
+    # Inside GPU context → CUDA is available
+    print("inside greet, torch.cuda.is_available():", torch.cuda.is_available())
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-demo = gr.Interface(fn=greet, inputs=gr.Number(), outputs=gr.Text())
-demo.launch()
+    # Move the tensor to the active device and add the input
+    t0 = zero.to(device)
+    tn = torch.tensor([float(n)], device=device)
+    out = t0 + tn
+
+    # Log where it ran; on ZeroGPU this should be cuda:0
+    print("tensor device during compute:", out.device)
+
+    # Return a friendly message
+    return f"Hello {out.item():.3f} (device: {out.device})"
+
+demo = gr.Interface(fn=greet, inputs=gr.Number(label="Add to zero"), outputs=gr.Text(label="Result"))
+
+if __name__ == "__main__":
+    # Minimal launch so the Space just builds
+    demo.launch()