it does not generate the same token or meaningful tokens with llama3.2-1B

[chakpongchung]

I tried to use the same prompt from this example and compare the output tokens with those from vllm with llama3.2-1B.
https://github.com/vllm-project/vllm/blob/main/examples/offline_inference.py#L5

It does not generate the expected outcome. To reproduce it, I use the following prompt and seed:

    prompt = "Hello, my name is"
    random.seed(32)
    torch.manual_seed(32)

    output_text = tokenizer.decode(output_ids)
    print(f'output_text: {output_text}') # output_text: ersetvod
    

Looks like it is missing a prefill step.