Didier/m2m100-12B-avg-5-ckpt

M2M100 12B (average of last 5 checkpoints)

• This is a copy of the model repository facebook/m2m100-12B-avg-5-ckpt, "a multilingual encoder-decoder (seq-to-seq) model trained for Many-to-Many multilingual translation".
• The model in the original repository is a single file of size 47.2 GB which can be an issue for people behind proxies where downloading files greater than xxGB is not permitted.

Steps:

• The model weights have been converted to bfloat16.
• The model file has been chunked into files no greater than 5 GB.

Usage

Sample usage:

from transformers import M2M100Tokenizer, M2M100ForConditionalGeneration
from threading import Lock

model_name = 'Didier/m2m100-12B-avg-5-ckpt'

device = 'mps' # if on Apple silicon
tokenizer = M2M100Tokenizer.from_pretrained(model_name)
model = M2M100ForConditionalGeneration.from_pretrained(
    model_name, device_map=device, low_cpu_mem_usage=True)
lock = Lock()

def translate(text: str, src_lang: str, tgt_lang: str) -> str:
    # Acquire lock to set src_lang and tokenize atomically
    with lock:
        tokenizer.src_lang = src_lang
        input_ids = tokenizer([text,], return_tensors="pt").input_ids.to(model.device)
    
    # Generate translation (outside the lock to allow parallel model 
    outputs = model.generate(
        input_ids=input_ids,
        forced_bos_token_id=tokenizer.get_lang_id(tgt_lang))
    translation = tokenizer.batch_decode(
        outputs, skip_special_tokens=True)[0]
    
    return translation


text = "ist der Ruf erst ruiniert, lebt es sich ganz ungeniert."
src_lang = 'de'
tgt_lang = 'en'

translation = translate(text, src_lang, tgt_lang)
print(f"{translation=}")

# --> "Once your reputation is ruined, you can live quite freely."