---
language:
- vi
- en
tags:
- translation
- pytorch
- custom-architecture
pipeline_tag: text-generation
base_model:
- "FacebookAI/xlm-roberta-large"
---

# DLM Vi2En

This is a Vietnamese to English translation model based on the DLM architecture.

Base model is: `"FacebookAI/xlm-roberta-large"`

## Requirements

Please ensure you have the following library versions installed:


```bash
pip install torch>=2.9.1 transformers>=4.57.3
````

## Inference

Below is the Python code to run the model. It automatically utilizes the GPU if available and loads the model from the local cache after the first run.

```python
import torch
from transformers import AutoTokenizer, AutoModel

# 1. Configuration
repo_id = "myduy/dlm-vi2en-checkpoint-90000"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# 2. Load Model & Tokenizer
# trust_remote_code=True is required for custom architectures
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True).to(device)
model.eval()

# 3. Prepare Input
text = "cậu có muốn đến nghe không?"
inputs = tokenizer(text, return_tensors="pt").to(device)

# 4. Generate
with torch.no_grad():
    output_tokens = model.generate(
        inputs.input_ids, 
        max_iterations=50, 
        temperature=1.0, 
        strategy="reparam-uncond-deterministic-cosine"
    )

# 5. Decode Output
output_text = tokenizer.batch_decode(output_tokens, skip_special_tokens=True)[0]

print(f"Input: {text}")
print(f"Output: {output_text}")
```