How do training_step wrappers affect backprop?

[OverLordGoldDragon]

def training_step(self, batch, batch_idx):
    try:
        loss = self._training_step(batch, batch_idx):
    except:
        loss = self._training_step(batch, batch_idx):

and

try:
    trainer.fit(model)
except:
    model.training_step(batch, batch_idx)

do not behave the same, despite having identical execution order. Latter no longer reproduces

RuntimeError: Function 'CudnnConvolutionBackward0' returned nan values in its 1th output.

while former keeps reproducing it no matter how many times _training_step is called. Code executes with torch.use_deterministic_algorithms(True) and same batch and batch_idx is passed in each case. Wrapper code appears to modify the model in some way.

What are some likely culprits?