Testing.assert_close gives different max difference than manual calculation

The following 2 snippets give different results for the same tensors:

testing.assert_close(a, b) # prints "Greatest absolute difference: 0.001220703125"
print(f"Max diff (flash): {(a - b).abs().max()}") # prints "Max diff (flash): 0.03125"

Can someone explain how this might be possible?

I get the same result:

a = torch.randn(10, 10)
b = a.clone()
b[0, 0] *= 0.1
torch.testing.assert_close(a, b)
# Mismatched elements: 1 / 100 (1.0%)
# Greatest absolute difference: 0.8026720210909843 at index (0, 0) (up to 1e-05 allowed)
# Greatest relative difference: 8.999999832919992 at index (0, 0) (up to 1.3e-06 allowed)

print((a - b).abs().max())
# tensor(0.8027)

so could you post a minimal, executable code snippet reproducing the issue, please?