autorev_inf.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
adapter_id = "Maitreya152/AutoRev"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
num_added = tokenizer.add_special_tokens({"pad_token": "<PAD>"})
tokenizer.padding_side = "right"

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16, 
    device_map="auto"           
)

if num_added > 0:
    base_model.resize_token_embeddings(len(tokenizer))
base_model.config.pad_token_id = tokenizer.pad_token_id

model = PeftModel.from_pretrained(base_model, adapter_id)

passages = """
"""

prompt = f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
Generate a structured feedback for the research paper passages provided below. The feedback should include a summary of the paper, its strengths, weaknesses, and questions for the authors. Consider that the feedback is being given for a paper submitted to the ICLR conference.
### Research Paper Passages:
{passages.strip()}
### Feedback for the paper:
"""

inputs = tokenizer(
    prompt, 
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=6000,         
    do_sample=True,
    temperature=0.7,            
    top_p=0.9,
    eos_token_id=tokenizer.eos_token_id
)

input_length = inputs.input_ids.shape[1]
generated_tokens = outputs[0][input_length:]
response = tokenizer.decode(generated_tokens, skip_special_tokens=True)

print(response)