ai-agents-causal-lm-past-key-v-when-using-past-key-values-with-padded-batches-in--7c12d962

THEN Manually specify correct position_ids for the new tokens, computed as the sequence lengths of each example before adding new tokens. Alternatively, use left-padding when tokenizing (tokenizer(padding_side='left')) so that padding does not disrupt position alignment.

```python
# Fix 1: manually provide position_ids
position_ids = torch.tensor([[3],[4]], dtype=torch.long)  # adjust per batch
outputs2 = model(input_ids=inputs2['input_ids'], attention_mask=attention_mask, past_key_values=outputs1.past_key_values, position_ids=position_ids)

# Fix 2: use left-padding
tokenizer = AutoTokenizer.from_pretrained('gpt2', padding_side='left')
# then tokenize and use past_key_values as usual
```