MPS torch.var cannot handle dim=0 on zero-dimensional input tensor #160738
Description
🕵️♂️ Detected with FACTO
🐛 Describe the bug
As a general rule, torch operations allow accessing dimension 0 on a zero-dimensional tensor (returning back the same tensor). In this case, the CPU torch.var operation can handle zero-dimensional input with dim=0, but MPS throws error:
import torch
def test_var(device):
x = torch.tensor(3.0, device=device)
try:
output = torch.var(x, dim=0)
print(f"var test succeeds for device: {device}. output: {output}")
except Exception as e:
print(f"var test fails for device: {device}: {e}")
test_var(device = "cpu")
test_var(device = "mps")terminal output:
var test succeeds for device: cpu. output: nan
var test fails for device: mps: var_mps: reduction dim must be in the range of input shape
Versions
nightly