Spaces:
Running
on
Zero
Running
on
Zero
File size: 2,275 Bytes
e1f2ca6 5e65e89 e1f2ca6 d2dc325 5e65e89 2e1c27e e1f2ca6 5e65e89 e1f2ca6 |
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 |
import random
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import spaces
# Available models
MODEL_OPTIONS_TRANSLATE = {
#"Flan-T5-A-Dhivehi-Latin Model": "alakxender/flan-t5-base-dhivehi-en-latin-v2",
"Flan-T5-B-Dhivehi-Latin Model": "alakxender/flan-t5-base-dhivehi-en-latin",
"MT5-B-Dhivehi-English Model": "alakxender/mt5-base-dv-en",
"MT5-B1-Dhivehi-English Model": "alakxender/mt5-base-dv-en-md",
}
# 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 = AutoModelForSeq2SeqLM.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]
max_input_length = 128
max_output_length = 128
@spaces.GPU()
def translate(instruction, input_text, model_choice, max_new_tokens=128, num_beams=4, repetition_penalty=1.2, no_repeat_ngram_size=3):
model_dir = MODEL_OPTIONS_TRANSLATE[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,
max_length=max_input_length
)
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_new_tokens,
"min_length":10,
"num_beams":num_beams,
"early_stopping":True,
"no_repeat_ngram_size":no_repeat_ngram_size,
"repetition_penalty":repetition_penalty,
"do_sample":False,
"pad_token_id":tokenizer.pad_token_id,
"eos_token_id":tokenizer.eos_token_id
}
with torch.no_grad():
outputs = model.generate(**gen_kwargs)
decoded_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
return decoded_output |