Cisco_DevNet_Sandbox_Running_Config

How to use this model

To use this model:

import torch
import transformers
import pyreft

device = "cuda"

# Load the base model
model_name_or_path = "meta-llama/Meta-Llama-3-8B"
model = transformers.AutoModelForCausalLM.from_pretrained(
    model_name_or_path, torch_dtype=torch.bfloat16, device_map={"": device}
)

# Load the ReFT model
reft_model = pyreft.ReftModel.load(
    "./CiscoDevNetSandboxRunningConfig", model
)

# Ensure the ReFT model components are also on the GPU
reft_model.set_device(device)

# Define the prompt template
prompt_no_input_template = """<s>[INST] <<SYS>>
You are a computer networking expert specialized in Cisco IOS XE running configurations.
<</SYS>>

%s [/INST]
"""

# Load the tokenizer
tokenizer = transformers.AutoTokenizer.from_pretrained(
    model_name_or_path, model_max_length=2048, 
    padding_side="right", use_fast=False
)

# Set pad_token as eos_token
tokenizer.pad_token = tokenizer.eos_token

def generate_response(instruction):
    # Tokenize and prepare the input
    prompt = prompt_no_input_template % instruction
    prompt = tokenizer(prompt, return_tensors="pt").to(device)
    
    base_unit_location = prompt["input_ids"].shape[-1] - 1  # Last position

    # Move all relevant tensors and operations to the GPU
    prompt = {key: value.to(device) for key, value in prompt.items()}  # Ensure the prompt is on the GPU
    
    # Generate the response using the reft_model
    _, reft_response = reft_model.generate(
        prompt, unit_locations={"sources->base": (None, [[[base_unit_location]]])},
        intervene_on_prompt=True, max_new_tokens=512, do_sample=True, 
        eos_token_id=tokenizer.eos_token_id, early_stopping=True
    )
    
    fine_tuned_answer = tokenizer.decode(reft_response[0], skip_special_tokens=True)
    return fine_tuned_answer

while True:
    instruction = input("You: ")
    
    if instruction.lower() == "exit":
        print("Goodbye!")
        break

    # Generate and print the response
    fine_tuned_answer = generate_response(instruction)
    print(f"Question: {instruction}")
    print(f"Fine-tuned Answer: {fine_tuned_answer}")

    # Log intervention details
    print(f"Intervention applied: {reft_model.interventions}")