app.py

"""Python file to serve as the frontend"""
import streamlit as st
from streamlit_chat import message

from langchain.chains import ConversationChain, LLMChain
from langchain import PromptTemplate
from langchain.llms.base import LLM
from langchain.memory import ConversationBufferWindowMemory
from typing import Optional, List, Mapping, Any

import torch
from peft import PeftModel
import transformers

from transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig
from transformers import BitsAndBytesConfig

tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf")



quantization_config = BitsAndBytesConfig(llm_int8_enable_fp32_cpu_offload=True)

model = LlamaForCausalLM.from_pretrained(
    "decapoda-research/llama-7b-hf",
    # load_in_8bit=True,
    # torch_dtype=torch.float16,
    device_map="auto",
    # device_map={"":"cpu"},
    max_memory={"cpu":"15GiB"}
    quantization_config=quantization_config
)
model = PeftModel.from_pretrained(
    model, "tloen/alpaca-lora-7b",
    # torch_dtype=torch.float16,
    device_map={"":"cpu"},
)

device = "cpu"
print("model device :", model.device, flush=True)
# model.to(device)
model.eval()

def evaluate_raw_prompt(
        prompt:str,
        temperature=0.1,
        top_p=0.75,
        top_k=40,
        num_beams=4,
        **kwargs,
):
    inputs = tokenizer(prompt, return_tensors="pt")
    input_ids = inputs["input_ids"].to(device)
    generation_config = GenerationConfig(
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        num_beams=num_beams,
        **kwargs,
    )
    with torch.no_grad():
        generation_output = model.generate(
            input_ids=input_ids,
            generation_config=generation_config,
            return_dict_in_generate=True,
            output_scores=True,
            max_new_tokens=256,
        )
    s = generation_output.sequences[0]
    output = tokenizer.decode(s)
    # return output
    return output.split("### Response:")[1].strip()

class AlpacaLLM(LLM):
    temperature: float
    top_p: float
    top_k: int
    num_beams: int
    @property
    def _llm_type(self) -> str:
        return "custom"
    
    def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
        if stop is not None:
            raise ValueError("stop kwargs are not permitted.")
        answer = evaluate_raw_prompt(prompt, 
                            top_p= self.top_p,
                            top_k= self.top_k,
                            num_beams= self.num_beams,
                            temperature= self.temperature
                            )
        return answer
    
    @property
    def _identifying_params(self) -> Mapping[str, Any]:
        """Get the identifying parameters."""
        return {
            "top_p": self.top_p,
            "top_k": self.top_k,
            "num_beams": self.num_beams,
            "temperature": self.temperature
            }


template = """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:
You are a chatbot, you should answer my last question very briefly. You are consistent and non repetitive. 
### Chat:
{history}
Human: {human_input}
### Response:"""

prompt = PromptTemplate(
    input_variables=["history","human_input"],
    template=template,
)


def load_chain():
    """Logic for loading the chain you want to use should go here."""
    llm = AlpacaLLM(top_p=0.75, top_k=40, num_beams=4, temperature=0.1)
    # chain = ConversationChain(llm=llm)
    chain = LLMChain(llm=llm, prompt=prompt, memory=ConversationBufferWindowMemory(k=2))
    return chain

chain = load_chain()

# # From here down is all the StreamLit UI.
# st.set_page_config(page_title="LangChain Demo", page_icon=":robot:")
# st.header("LangChain Demo")

# if "generated" not in st.session_state:
#     st.session_state["generated"] = []

# if "past" not in st.session_state:
#     st.session_state["past"] = []


# def get_text():
#     input_text = st.text_input("Human: ", "Hello, how are you?", key="input")
#     return input_text


# user_input = get_text()

# if user_input:
#     output = chain.predict(human_input=user_input)

#     st.session_state.past.append(user_input)
#     st.session_state.generated.append(output)

# if st.session_state["generated"]:

#     for i in range(len(st.session_state["generated"]) - 1, -1, -1):
#         message(st.session_state["generated"][i], key=str(i))
#         message(st.session_state["past"][i], is_user=True, key=str(i) + "_user")

st.title("ChatAlpaca")

if "history" not in st.session_state:
    st.session_state.history = []
    st.session_state.history.append({"message": "Hey, I'm a Alpaca chatBot. Ask whatever you want!", "is_user": False})
    

def generate_answer():
    user_message = st.session_state.input_text
    inputs = tokenizer(st.session_state.input_text, return_tensors="pt")
    result = model.generate(**inputs)
    message_bot = tokenizer.decode(result[0], skip_special_tokens=True)  # .replace("<s>", "").replace("</s>", "")
    st.session_state.history.append({"message": user_message, "is_user": True})
    st.session_state.history.append({"message": message_bot, "is_user": False})
    
st.text_input("Response", key="input_text", on_change=generate_answer)

for chat in st.session_state.history:
    st_message(**chat)