Bug: golden_run() modifies original model behavior, breaking observers and baselines

[bcarrpe]

fi_model.golden_run() produces different outputs than the original unwrapped model, causing incorrect observer metrics (RMSE, MAE, EqualRate) and baseline measurements.

To reproduce the issue:

import torch
from torchvision.models import vgg11, VGG11_Weights
from mrfi import MRFI, EasyConfig

model = vgg11(weights=VGG11_Weights.DEFAULT).eval()
x = torch.randn(1, 3, 224, 224)

# Original model output
y_original = model(x)

# MRFI golden run output
fi_model = MRFI(model, EasyConfig.load_preset('default_fi'))
with fi_model.golden_run():
    y_golden = fi_model(x)

# They should be identical but aren't
print(torch.abs(y_original - y_golden).max())  # Non-zero!

golden_run() disables fault injection but leaves model hooks active, intercepting and processing data.

[fffasttime]

Thanks for testing! However, this is actually expected behavior.

The default_fi configuration includes model quantization, and the "golden run" mode is designed to disable fault injection while retaining quantization. The difference you observed between y_original and y_golden is due to quantization error. When comparing the impact of fault injection on quantized models, we generally use the quantized model as the golden run baseline, and do not treat quantization-induced differences as faults.

If you'd like to compare against the pre-quantization model, it's better to use y_original. Alternatively, you can use a preset like float_fi, which excludes quantization—the golden run result will then remain consistent.