app.py

import os
import spaces
import gradio as gr
import torch
torch.jit.script = lambda f: f  # Avoid script error in lambda

# Initialize non-GPU components first
from langchain.embeddings import HuggingFaceInstructEmbeddings
from langchain.vectorstores import Chroma
from langchain.prompts import PromptTemplate
from langchain.chains import RetrievalQA

# System prompts
DEFAULT_SYSTEM_PROMPT = """
Based on the information in this document provided in context, answer the question as accurately as possible in 1 or 2 lines. If the information is not in the context,
respond with "I don't know" or a similar acknowledgment that the answer is not available.
""".strip()

SYSTEM_PROMPT = "Use the following pieces of context to answer the question at the end. Do not provide commentary or elaboration more than 1 or 2 lines.?"

def generate_prompt(prompt: str, system_prompt: str = DEFAULT_SYSTEM_PROMPT) -> str:
    return f"""
[INST] <<SYS>>
{system_prompt}
<</SYS>>

{prompt} [/INST]
""".strip()

template = generate_prompt(
    """
{context}

Question: {question}
""",
    system_prompt=SYSTEM_PROMPT,
)

prompt_template = PromptTemplate(template=template, input_variables=["context", "question"])

# Initialize database and embeddings
embeddings = HuggingFaceInstructEmbeddings(
    model_name="hkunlp/instructor-base",
    model_kwargs={"device": "cpu"}
)

db = Chroma(
    persist_directory="db",
    embedding_function=embeddings
)

def initialize_model():
    from transformers import AutoTokenizer, TextStreamer, pipeline, AutoModelForCausalLM
    from langchain.llms import HuggingFacePipeline
    
    model_id = "meta-llama/Llama-3.2-3B-Instruct"
    token = os.environ.get("HF_TOKEN")
    
    tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        token=token,
    )
    
    if torch.cuda.is_available():
        model = model.to("cuda")
    
    return model, tokenizer

@spaces.GPU
def respond(message, history, system_message, max_tokens, temperature, top_p):
    try:
        # Initialize model components inside GPU context
        model, tokenizer = initialize_model()
        from transformers import TextStreamer, pipeline
        
        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
        text_pipeline = pipeline(
            "text-generation",
            model=model,
            tokenizer=tokenizer,
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            repetition_penalty=1.15,
            streamer=streamer,
        )
        
        llm = HuggingFacePipeline(pipeline=text_pipeline)
        qa_chain = RetrievalQA.from_chain_type(
            llm=llm,
            chain_type="stuff",
            retriever=db.as_retriever(search_kwargs={"k": 2}),
            return_source_documents=False,
            chain_type_kwargs={"prompt": prompt_template}
        )
        
        response = qa_chain.invoke({"query": message})
        yield response["result"]
        
    except Exception as e:
        yield f"An error occurred: {str(e)}"

# Create Gradio interface
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(
            value=DEFAULT_SYSTEM_PROMPT,
            label="System Message",
            lines=3,
            visible=False
        ),
        gr.Slider(
            minimum=1,
            maximum=2048,
            value=500,
            step=1,
            label="Max new tokens"
        ),
        gr.Slider(
            minimum=0.1,
            maximum=4.0,
            value=0.1,
            step=0.1,
            label="Temperature"
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)"
        ),
    ],
    title="ROS2 Expert Assistant",
    description="Ask questions about ROS2, navigation, and robotics. I'll provide concise answers based on the available documentation.",
)

if __name__ == "__main__":
    demo.launch()