File size: 1,697 Bytes
7f028d4
 
 
 
 
024e1b5
7f028d4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f8636f6
7f028d4
 
 
 
 
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
import torch
import gradio as gr
from transformers import GPT2LMHeadModel, GPT2TokenizerFast

# Replace with your HF username and repo name
MODEL_REPO = "i3-lab/i3-GPT2" 

# Load model and tokenizer
tokenizer = GPT2TokenizerFast.from_pretrained(MODEL_REPO)
model = GPT2LMHeadModel.from_pretrained(MODEL_REPO)

# Move to GPU if the Space has one, else CPU
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

def generate_response(message, history):
    # Construct the prompt using the same format as your training script
    prompt = ""
    for user_msg, assistant_msg in history:
        prompt += f"User: {user_msg}\nAssistant: {assistant_msg}<|endoftext|>\n"
    prompt += f"User: {message}\nAssistant:"

    # Tokenize input
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    
    # Generate
    with torch.no_grad():
        output_tokens = model.generate(
            **inputs,
            max_new_tokens=150,
            do_sample=True,
            top_p=0.9,
            temperature=0.7,
            pad_token_id=tokenizer.eos_token_id,
            repetition_penalty=1.2
        )
    
    # Extract only the newly generated text
    response = tokenizer.decode(output_tokens[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
    
    # Clean up formatting (cutting off if the model generates a new 'User:' tag)
    clean_response = response.split("User:")[0].strip()
    return clean_response

# Launch Gradio Chat Interface
demo = gr.ChatInterface(
    fn=generate_response,
    title="i3-GPT",
    examples=["Tell me a joke.", "What is the capital of France?", "How does a lightbulb work?"]
)

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