__bool__ (and data dependent control flow)

HF BERT data-dependent control flow:

if self.config.pad_token_id in input_ids[:, [-1, 0]]:
   4348     warn_string = (
   4349         "We strongly recommend passing in an `attention_mask` since your input_ids may be padded. See "
   4350         "https://huggingface.co/docs/transformers/troubleshooting"
   4351         "#incorrect-output-when-padding-tokens-arent-masked."
   4352     )
   4354     # If the pad token is equal to either BOS, EOS, or SEP, we do not know whether the user should use an
   4355     # attention_mask or not. In this case, we should still show a warning because this is a rare case.

input_ids is a tensor, that ultimately makes us fail on __bool__ for tensors.

Repro:

import torch, thunder, transformers

m = transformers.BertForSequenceClassification(transformers.BertConfig())
jm = thunder.jit(m)
a = torch.randint(1, 20, (1, 25))
jm(a)

Start by running the provided Python repro with PyTorch, Thunder, and the BERT model, then trace the tensor bool failure during data-dependent control flow. Done means the repro compiles and invokes BertForSequenceClassification without failing on tensor truth-value evaluation.