File size: 2,187 Bytes
cd9f86b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import random
import numpy as np
import torch
from transformers import AutoTokenizer, T5ForConditionalGeneration
import spaces


# Available models
MODEL_OPTIONS_INSTRUCT_TUNED = {
    "EXT1 Model": "alakxender/flan-t5-base-alpaca-dv-ext"
}

# Cache for loaded models/tokenizers
MODEL_CACHE = {}

def get_model_and_tokenizer(model_dir):
    if model_dir not in MODEL_CACHE:
        print(f"Loading model: {model_dir}")
        tokenizer = AutoTokenizer.from_pretrained(model_dir)
        model = T5ForConditionalGeneration.from_pretrained(model_dir)
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        print(f"Moving model to device: {device}")
        model.to(device)
        MODEL_CACHE[model_dir] = (tokenizer, model)
    return MODEL_CACHE[model_dir]

@spaces.GPU()
def generate_response_tuned(instruction, input_text, seed, model_choice,max_tokens, temperature, num_beams):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed) 

    model_dir = MODEL_OPTIONS_INSTRUCT_TUNED[model_choice]
    tokenizer, model = get_model_and_tokenizer(model_dir)

    combined_input = f"{instruction.strip()} {input_text.strip()}" if input_text else instruction.strip()
    inputs = tokenizer(
        combined_input,
        return_tensors="pt",
        truncation=True
    )
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    inputs = {k: v.to(device) for k, v in inputs.items()}

    gen_kwargs = {
        **inputs,
        "max_length":max_tokens,
        "num_beams":num_beams,
        "do_sample":(temperature > 0.0),
        "temperature":temperature,
        "repetition_penalty":1.2,
        "top_p":0.95,
        "top_k":50
    }

    with torch.no_grad():
        outputs = model.generate(**gen_kwargs)
    decoded_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Trim to the last period
    if '.' in decoded_output:
        last_period = decoded_output.rfind('.')
        decoded_output = decoded_output[:last_period+1]

    decoded_output = ' '.join(decoded_output.split())
    return decoded_output