llama2-call-summarization

install transformers, peft, accelerate & BitsAndBytes

import transformers
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from torch import cuda, bfloat16

base_model_id = 'meta-llama/Llama-2-7b-chat-hf'

device = f'cuda:{cuda.current_device()}' if cuda.is_available() else 'cpu'

bnb_config = transformers.BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type='nf4',
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=bfloat16
)


hf_auth = "your-hf-access-token"
model_config = transformers.AutoConfig.from_pretrained(
    base_model_id,
    use_auth_token=hf_auth
)

model = transformers.AutoModelForCausalLM.from_pretrained(
    base_model_id,
    trust_remote_code=True,
    config=model_config,
    quantization_config=bnb_config,
    device_map='auto',
    use_auth_token=hf_auth
)

config = PeftConfig.from_pretrained("Ashishkr/llama2-call-summarization")
model = PeftModel.from_pretrained(model, "Ashishkr/llama2-call-summarization").to(device)

model.eval()
print(f"Model loaded on {device}")

tokenizer = transformers.AutoTokenizer.from_pretrained(
    base_model_id,
    use_auth_token=hf_auth
)

def llama_generate(
    model: AutoModelForCausalLM,
    tokenizer: AutoTokenizer,
    prompt: str,
    max_new_tokens: int = 128,
    temperature: float = 0.92):

    inputs = tokenizer(
        [prompt],
        return_tensors="pt",
        return_token_type_ids=False,
    ).to(
        device
    )

    # Check if bfloat16 is supported, otherwise use float16
    dtype_to_use = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16

    with torch.autocast("cuda", dtype=dtype_to_use):
        response = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            return_dict_in_generate=True,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id,
        )

    decoded_output = tokenizer.decode(
        response["sequences"][0],
        skip_special_tokens=True,
    )

    return decoded_output[len(prompt) :]

prompt = """
 instruction: "summarize this conversation :" \n

input: "Oli: I've talked to some people from the third year
Jacob: About the statistics exam?
Marcia: What did they say? 
Oli: Yeah, about the exam Oli: We need to prepare for a battle
Jacob: So it will be difficult 
Oli: They said it was the hardest exam ever 
Marcia: 😱
Oli: The questions were displayed on the screen 
Oli: One minute per question and it disappears 
Oli: They won't come back so if you didn't get your answer you're fucked 
Marcia: So we need to make the calculations really fast 
Jacob: That's insane
Oli: I know 
Oli: Very stressful 
Marcia: How are we even supposed to study for it?
Marcia: With a timer?
Oli: I guess
Marcia: Did anybody pass it last year
Oli: Some people did, but the majority had to take the second or even the third chance"\n

response:  """
response = llama_generate(
    model,
    tokenizer,
    prompt,
    max_new_tokens=100,
    temperature=0.9,
).split("<eos>")[0].strip()

print(response)