File size: 2,435 Bytes
06696b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
import gradio as gr
import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
# from retriever.vectordb_rerank import search_documents  # 🧠 RAG 검색기 뢈러였기
from services.rag_pipeline import rag_pipeline

model_name = "dasomaru/gemma-3-4bit-it-demo"


# 1. λͺ¨λΈ/ν† ν¬λ‚˜μ΄μ € 1회 λ‘œλ”©
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# πŸš€ model은 CPU둜만 λ¨Όμ € 올림 (GPU 아직 μ—†μŒ)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,  # 4bit modelμ΄λ‹ˆκΉŒ
    device_map="auto",  # βœ… μ€‘μš”: μžλ™μœΌλ‘œ GPU ν• λ‹Ή
    trust_remote_code=True,
)

# 2. μΊμ‹œ 관리
search_cache = {}

@spaces.GPU(duration=300)
def generate_response(query: str):
    tokenizer = AutoTokenizer.from_pretrained(
        "dasomaru/gemma-3-4bit-it-demo",
        trust_remote_code=True,
        )
    model = AutoModelForCausalLM.from_pretrained(
        "dasomaru/gemma-3-4bit-it-demo",
        torch_dtype=torch.float16,  # 4bit modelμ΄λ‹ˆκΉŒ
        device_map="auto",  # βœ… μ€‘μš”: μžλ™μœΌλ‘œ GPU ν• λ‹Ή
        trust_remote_code=True,

        )
    model.to("cuda")    

    if query in search_cache:
        print(f"⚑ μΊμ‹œ μ‚¬μš©: '{query}'")
        return search_cache[query]

    # πŸ”₯ rag_pipeline을 ν˜ΈμΆœν•΄μ„œ 검색 + 생성
    # 검색
    top_k = 5
    results = rag_pipeline(query, top_k=top_k)

    # κ²°κ³Όκ°€ list일 경우 ν•©μΉ˜κΈ°
    if isinstance(results, list):
        results = "\n\n".join(results)

    search_cache[query] = results
    # return results

    inputs = tokenizer(results, return_tensors="pt").to(model.device)  # βœ… model.device
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        top_k=50,
        do_sample=True,
    )

    return tokenizer.decode(outputs[0], skip_special_tokens=True)

    
# 3. Gradio μΈν„°νŽ˜μ΄μŠ€
demo = gr.Interface(
    fn=generate_response,
    # inputs=gr.Textbox(lines=2, placeholder="μ§ˆλ¬Έμ„ μž…λ ₯ν•˜μ„Έμš”"),
    inputs="text",
    outputs="text",
    title="Law RAG Assistant",
    description="법령 기반 RAG νŒŒμ΄ν”„λΌμΈ ν…ŒμŠ€νŠΈ",
)

# demo.launch(server_name="0.0.0.0", server_port=7860)  # πŸš€ API 배포 μ€€λΉ„ κ°€λŠ₯
# demo.launch()
demo.launch(debug=True)