m2m100_418M_ewe_fr_rel_news_ft-onnx

ONNX export of masakhane/m2m100_418M_ewe_fr_rel_news_ft.

Translation direction: Ewe (ewe) -> French (fr).

Attribution

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_ewe_fr_rel_news_ft \
  --task text2text-generation-with-past --no-post-process <outdir>

The int8 folder was made with onnxruntime.quantization.quantize_dynamic (QuantType.QInt8, EnableSubgraph=True).

Files

Path	Precision	Size
*.onnx (root)	fp32	4525 MB
int8/*.onnx	int8 dynamic	1156 MB

Encoder, decoder and decoder-with-past are separate graphs (--no-post-process); the merged decoder is not produced because merging exhausts memory on this model.

Parity

8 sentences from the MAFAND-MT fr-ewe test split (source: masakhane-io/lafand-mt, data/json_files/fr-ewe/test.json), PyTorch original vs ONNX, exact string match:

• fp32, num_beams=1 (greedy): 100% (8/8)
• fp32, num_beams=4: 100% (8/8)
• int8, num_beams=1 (greedy): 25% (2/8, reported, not gated)
• int8, num_beams=4: 50% (4/8, reported, not gated)

int8 divergences are topically coherent paraphrases, not garbage output (see the export script's parity log); the fp32 graph is the one this repository gates on.

Selecting the language

This is a single-direction model. The target language is already fixed in generation_config.json via forced_bos_token_id (__fr__, real French), so you do not set it yourself. You must set the source language on the tokenizer:

tokenizer.src_lang = "sw"

Masakhane reused the existing M2M100 __sw__ (Swahili) token as a stand-in for Ewe, which M2M100 does not cover - the same stand-in the reverse-direction fr-ewe model uses for its Ewe target. Using a different value silently degrades output.

Usage

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("TigreGotico/m2m100_418M_ewe_fr_rel_news_ft-onnx")
tok.src_lang = "sw"
model = ORTModelForSeq2SeqLM.from_pretrained("TigreGotico/m2m100_418M_ewe_fr_rel_news_ft-onnx")

enc = tok("Le Côte d'Ivoire: Ahmed si nye atiglinyi ahanola la ga dze kpɔ, eye wòle nu gblem le yiyim.", return_tensors="pt")
out = model.generate(**enc, num_beams=4, max_new_tokens=64)
print(tok.batch_decode(out, skip_special_tokens=True)[0])
# -> Côte d'Ivoire : Ahmed, l'homme alcoolique d'un éléphant, a déjà fait ses preuves et il est en train de se détériorer

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