m2m100_418M_bam_fr_rel_news_ft-onnx

ONNX export of masakhane/m2m100_418M_bam_fr_rel_news_ft.

Translation direction: Bambara (bam) -> 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_bam_fr_rel_news_ft \
  --task text2text-generation-with-past --no-post-process <outdir>

Files

Path	Precision	Size
*.onnx (root)	fp32	4754 MB
int8/*.onnx	int8 dynamic	1201 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-bam test split, PyTorch original vs ONNX, num_beams=4, max_new_tokens=64, exact string match:

• fp32: 100%
• int8: 25% (reported, not gated)

Selecting the language

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

tokenizer.src_lang = "sw"

Masakhane reused existing M2M100 language tokens as stand-ins for languages M2M100 does not cover, so src_lang is "sw" here, which is not the ISO code of Bambara. Using a different value silently degrades output.

Usage

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("TigreGotico/m2m100_418M_bam_fr_rel_news_ft-onnx")
tok.src_lang = "sw"
model = ORTModelForSeq2SeqLM.from_pretrained("TigreGotico/m2m100_418M_bam_fr_rel_news_ft-onnx")

enc = tok("Minisiriɲɛmɔgɔ ye nin ko kɛ sababu ye ka Jamana Labɛn kura Lajɛbaw nata kɛli bange kunnafondilaw ye.", return_tensors="pt")
out = model.generate(**enc, num_beams=4, max_new_tokens=64)
print(tok.batch_decode(out, skip_special_tokens=True)[0])
# -> Le Premier ministre a saisi l’occasion pour engendrer la presse sur la prochaine conférence des Assemblées nationales de réhabilitation.

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