File size: 2,827 Bytes
816ccb1
 
 
 
9d0df30
816ccb1
 
9d0df30
 
816ccb1
 
 
9d0df30
 
 
 
 
 
 
 
 
 
816ccb1
 
451c534
 
9d0df30
816ccb1
 
 
 
 
 
 
9d0df30
27e057c
451c534
9d0df30
451c534
 
 
9d0df30
 
 
 
 
 
 
 
 
 
 
451c534
9d0df30
27e057c
9d0df30
27e057c
816ccb1
9d0df30
816ccb1
 
 
 
 
 
 
 
27e057c
 
 
816ccb1
 
 
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
import gradio as gr
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Set up device: use GPU if available, else CPU.
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load the fine-tuned model and tokenizer.
model_name = "aarohanverma/text2sql-flan-t5-base-qlora-finetuned"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to(device)
tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-base")

# For CPU inference, convert the model to FP32 for better compatibility.
if device.type == "cpu":
    model = model.float()

# Optionally compile the model for speed improvements (requires PyTorch 2.0+).
try:
    model = torch.compile(model)
except Exception as e:
    print("torch.compile optimization failed:", e)

def generate_sql(context: str, query: str) -> str:
    """
    Generates a SQL query given the provided context and natural language query.
    Constructs a prompt from the inputs, then performs deterministic generation
    using beam search with repetition handling.
    """
    prompt = f"""Context:
{context}
Query:
{query}
Response:
"""
    # Tokenize the prompt with truncation and max length; move to device.
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512).to(device)
    
    # Ensure the decoder start token is set.
    if model.config.decoder_start_token_id is None:
        model.config.decoder_start_token_id = tokenizer.pad_token_id

    # Generate SQL output with no_grad to optimize CPU usage.
    with torch.no_grad():
        generated_ids = model.generate(
            input_ids=inputs["input_ids"],
            decoder_start_token_id=model.config.decoder_start_token_id,
            max_new_tokens=100,
            temperature=0.0,         # Deterministic output
            num_beams=5,
            repetition_penalty=1.2,
            early_stopping=True,
        )
    
    # Decode and clean the generated SQL statement.
    generated_sql = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
    generated_sql = generated_sql.split(";")[0].strip() + ";"  # Keep only the first valid SQL query
    return generated_sql

# Create Gradio interface with two input boxes: one for context and one for query.
iface = gr.Interface(
    fn=generate_sql,
    inputs=[
        gr.Textbox(lines=8, label="Context", placeholder="Enter table schema, sample data, etc."),
        gr.Textbox(lines=2, label="Query", placeholder="Enter your natural language query here...")
    ],
    outputs="text",
    title="Text-to-SQL Generator",
    description="Enter your own context (e.g., database schema and sample data) and a natural language query. The model will generate the corresponding SQL statement.",
    theme="compact",
    allow_flagging="never"
)

iface.launch()