opus-mt-sk-en-onnx
ONNX export (fp32 + dynamic int8 quantized) of Helsinki-NLP/opus-mt-sk-en, a Marian 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-sk-en --task text2text-generation-with-past <out>
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: ~1178.0 MB | int8 size: ~547.6 MB
Parity check
Compared PyTorch (MarianMTModel) vs ONNX fp32 (ORTModelForSeq2SeqLM) on 2 sentences, greedy and beam=4 (max_new_tokens=64). Overall: greedy PASS, beam4 PASS.
- src: Toto je kniha.
- pytorch greedy: This is a book.
- onnx fp32 greedy: This is a book. (match)
- pytorch beam4: This is a book.
- onnx fp32 beam4: This is a book. (match)
- onnx int8 greedy: This is a book.
- src: Učím sa slovenčinu.
- pytorch greedy: I'm learning English.
- onnx fp32 greedy: I'm learning English. (match)
- pytorch beam4: I'm learning English.
- onnx fp32 beam4: I'm learning English. (match)
- onnx int8 greedy: I'm learning English.
Usage
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer
repo = "TigreGotico/opus-mt-sk-en-onnx"
tok = AutoTokenizer.from_pretrained(repo)
model = ORTModelForSeq2SeqLM.from_pretrained(repo) # fp32; pass subfolder="int8" for the quantized graphs
inputs = tok("Toto je kniha.", return_tensors="pt")
out = model.generate(**inputs, num_beams=4, max_new_tokens=64)
print(tok.decode(out[0], skip_special_tokens=True))
Exported for the OVOS / TigreGotico offline translation stack.