opus-mt-es-ca-onnx

opus-mt-es-ca-onnx

ONNX export (fp32 + dynamic int8 quantized) of Helsinki-NLP/opus-mt-es-ca, a Marian (es -> ca) translation model from the Helsinki-NLP OPUS-MT project.

License: apache-2.0 (inherited from the base model; verify at the source link above).

Export

optimum-cli export onnx --model Helsinki-NLP/opus-mt-es-ca --task text2text-generation-with-past /home/miro/tmp/claude-1000/-home-miro/849a3674-94c9-4aa5-b1f9-e596235b0a36/scratchpad/opusmt-onnx/out/es-ca

Quantized to int8 with onnxruntime.quantization.quantize_dynamic (QUInt8 weights).

File layout

./                      fp32 ONNX graphs (encoder_model.onnx, decoder_model.onnx, decoder_with_past_model.onnx) + tokenizer files
./int8/                 int8 dynamic-quantized ONNX graphs

fp32 size: ~763 MB | int8 size: ~341 MB

Sanity check

Compared PyTorch (MarianMTModel) vs ONNX fp32 (ORTModelForSeq2SeqLM, num_beams=4, max_new_tokens=48) on 3 sentences. Match: 3/3.

• src: Hola, ¿cómo estás hoy?
  • pytorch: Hola, com estàs avui?
  • onnx fp32: Hola, com estàs avui? (match)
• src: El tiempo está agradable esta mañana.
  • pytorch: El temps està agradable aquest matí.
  • onnx fp32: El temps està agradable aquest matí. (match)
• src: Me gustaría pedir un café, por favor.
  • pytorch: M'agradaria demanar un cafè, si us plau.
  • onnx fp32: M'agradaria demanar un cafè, si us plau. (match)

int8 sample output (not gated, for reference only):

• Hola, com estàs avui?

Usage

from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer

repo = "TigreGotico/opus-mt-es-ca-onnx"
tok = AutoTokenizer.from_pretrained(repo)
model = ORTModelForSeq2SeqLM.from_pretrained(repo)  # fp32; pass subfolder="int8" for the quantized graphs
inputs = tok("Hello, how are you?", return_tensors="pt")
out = model.generate(**inputs, num_beams=4, max_new_tokens=48)
print(tok.decode(out[0], skip_special_tokens=True))

Exported for the OVOS / TigreGotico offline translation stack.