fix llama runner (#6256)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/6256

imported-using-ghimport

Test Plan:
Imported from OSS

Run the following command and make sure it generates the right result:
```
python -m examples.models.llama2.runner.eager \
          --checkpoint /home/lunwenh/models/1B/consolidated.00.pth \
          --params /home/lunwenh/models/1B/params.json \
          --max_len 128 \
          --tokenizer /home/lunwenh/models/1B/tokenizer.model \
          --prompt "<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a good assistant<|eot_id|><|start_header_id|>user<|end_header_id|>

What is the capital of France?<|eot_id|><|start_header_id|>assistant<|end_header_id|>"
```
```
Response:
The capital of France is Paris.
 Tokens:
 [791, 6864, 315, 9822, 374, 12366, 13]
```

Reviewed By: mergennachin

Differential Revision: D64442223

Pulled By: helunwencser

fbshipit-source-id: 99ee56f73e472a7243b8896a35ee092f287edf6b