The original model was trained by Masakhane (masakhane-io/lafand-mt) as part of MAFAND-MT (Adelani et al., NAACL 2022, "A Few Thousand Translations Go a Long Way!"). It is a fine-tune of facebook/m2m100_418M on JW300 plus the MAFAND news corpus. Licence: AFL-3.0, as declared by the original repository. This repository only converts the weights to ONNX. All credit for the model belongs to Masakhane.

Export

optimum-cli export onnx --model masakhane/m2m100_418M_en_hau_rel_news_ft \
  --task text2text-generation-with-past --no-post-process <outdir>

Files

Path	Precision	Size
`*.onnx` (root)	fp32	4.6 GB
`int8/*.onnx`	int8 dynamic	3.0 GB

Encoder, decoder and decoder-with-past are separate graphs (--no-post-process).

Parity

10 general-domain English sentences, PyTorch original vs ONNX, exact string match against AutoModelForSeq2SeqLM.generate():

yaml

parity:
  fp32_greedy: 1.00   # 10/10
  fp32_beam4:  1.00   # 10/10
  int8_greedy: 0.50   # 5/10
  int8_beam4:  0.60   # 6/10

Decoding	fp32	int8
greedy	100% (10/10)	50% (5/10)
beam=4	100% (10/10)	60% (6/10)

fp32 is an exact reproduction of the original model. int8 dynamic quantization causes real quality loss on this checkpoint - prefer fp32, use int8 only where size is the binding constraint.

Selecting the language

Hausa has a native M2M100 language code (ha, unlike some of the other Masakhane fine-tunes that reuse unrelated codes as stand-ins). The target is already fixed via forced_bos_token_id in generation_config.json. You must still set the source language on the tokenizer:

tokenizer.src_lang = "en"

Usage

python

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("TigreGotico/m2m100_418M_en_hau_rel_news_ft-onnx")
tok.src_lang = "en"
model = ORTModelForSeq2SeqLM.from_pretrained("TigreGotico/m2m100_418M_en_hau_rel_news_ft-onnx", use_cache=True, use_merged=False)

enc = tok("The weather is very nice today.", return_tensors="pt")
out = model.generate(**enc, num_beams=4, max_new_tokens=64)
print(tok.batch_decode(out, skip_special_tokens=True)[0])
# -> A yau dai yanayi yana da kyau.

For the int8 build, pass subfolder="int8".