---
license: apache-2.0
base_model: Helsinki-NLP/opus-mt-hy-en
base_model_relation: quantized
tags:
  - onnx
  - translation
  - marian
library_name: transformers
pipeline_tag: translation
language:
  - hy
  - en
---

# opus-mt-hy-en-onnx

ONNX export (fp32 + dynamic int8) of [`Helsinki-NLP/opus-mt-hy-en`](https://huggingface.co/Helsinki-NLP/opus-mt-hy-en),
a Marian (hy -> en) translation model from the Helsinki-NLP OPUS-MT project.

**License:** apache-2.0, inherited unchanged from the base model.

## Contents

| Files | What it is |
|---|---|
| `encoder_model.onnx`, `decoder_model.onnx`, `decoder_with_past_model.onnx` | ONNX, float32 |
| `int8/` | same three graphs, dynamic int8 (`QInt8`, `MatMul` only, `/lm_head/MatMul` excluded) |
| `source.spm`, `target.spm`, `vocab.json` | `MarianTokenizer` over the original sentencepiece models |

fp32 size: 868 MB (three graphs) | int8 size: 679 MB

## Usage

```python
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer

repo = "TigreGotico/opus-mt-hy-en-onnx"
tok = AutoTokenizer.from_pretrained(repo)
model = ORTModelForSeq2SeqLM.from_pretrained(repo, use_cache=True, use_merged=False)  # fp32
# int8: ORTModelForSeq2SeqLM.from_pretrained(repo, subfolder="int8", use_cache=True, use_merged=False)
inputs = tok("Այսօր եղանակը շատ գեղեցիկ է:", return_tensors="pt")
out = model.generate(**inputs, num_beams=4, max_new_tokens=64)
print(tok.decode(out[0], skip_special_tokens=True))
```

## Parity with the original PyTorch model

10 general-domain **Armenian** sentences (source language for this hy->en
pair), exact-string-match of generated output against `MarianMTModel.generate()`
on the original `Helsinki-NLP/opus-mt-hy-en` checkpoint.

```yaml
parity:
  fp32_greedy: 1.00   # 10/10
  fp32_beam4:  1.00   # 10/10
  int8_greedy: 1.00   # 10/10
  int8_beam4:  0.90   # 9/10
```

| Decoding | fp32 exact match | int8 exact match |
|---|---|---|
| greedy (num_beams=1) | 10/10 (100.0%) | 10/10 (100.0%) |
| beam=4 | 10/10 (100.0%) | 9/10 (90.0%) |

fp32 is an exact reproduction of the original model at both decoding
settings, and int8 matches it almost everywhere - the best int8 agreement
observed among the Armenian/Azerbaijani models converted in this batch. The
single beam=4 disagreement was inspected by hand: ref "Can you help me find
the next train station?" vs int8 "Can you help me find my nearby railway
station?" - a paraphrase, not a degradation.

An earlier version of this card reported int8 80%/90% from a harness bug -
the parity script had reused a fixed **English** sentence set across all
opus-mt pairs, but this model's source language is Armenian, not English,
so that measurement compared the model against out-of-distribution input
and was discarded. The numbers above are the corrected, re-measured values
using real Armenian source sentences.