vLLM/Recipes
Preferred Networks

pfnet/plamo-2-translate

Post-trained PLaMo 2 translation model specialized for English/Japanese translation tasks.

PLaMo 2 post-trained for English/Japanese translation

View on HuggingFace
dense9.5B8,192 ctxvLLM 0.8.5+text
Guide

Overview

PLaMo 2 Translate is a specialized translation model developed by Preferred Networks for English/Japanese translation tasks. It is the post-trained main model in the PLaMo Translation Model family.

The family includes three checkpoints sharing the same architecture:

  • pfnet/plamo-2-translate: post-trained model for production translation
  • pfnet/plamo-2-translate-base: base checkpoint for fine-tuning experiments
  • pfnet/plamo-2-translate-eval: pair-wise evaluator that picks the better of two candidate translations

All three checkpoints can be served with the same vLLM flags shown here; just swap the model id.

Prerequisites

  • Hardware: 1x GPU with at least 24 GB VRAM, such as L40S, A30, or RTX 4090
  • vLLM: >= 0.8.5
  • License: PLaMo community license

Install vLLM

uv venv
source .venv/bin/activate
uv pip install -U vllm --torch-backend=auto

Launching the Server

vllm serve pfnet/plamo-2-translate \
  --trust-remote-code \
  --max-model-len 8192

For tighter memory budgets, reduce --max-model-len (the model card examples use 2000).

Prompt Format

Use the PLaMo operation-token structure. Place the source text after the input line and let generation continue after the output line.

<|plamo:op|>dataset
translation
<|plamo:op|>input lang=English
The text to translate goes here.
<|plamo:op|>output lang=Japanese

Client Usage

The <|plamo:op|> stop sequence prevents the model from generating beyond the translated text.

from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")
prompt = """<|plamo:op|>dataset
translation
<|plamo:op|>input lang=English
The text to translate goes here.
<|plamo:op|>output lang=Japanese
"""
resp = client.completions.create(
    model="pfnet/plamo-2-translate",
    prompt=prompt,
    max_tokens=1024,
    temperature=0.0,
    stop=["<|plamo:op|>"],
)
print(resp.choices[0].text)

Troubleshooting

  • Out of memory: lower --max-model-len to around 2000, or add --gpu-memory-utilization 0.95.
  • Garbled or empty output: check that the prompt uses the exact <|plamo:op|> operation-token format shown above.
  • Evaluator usage: serve pfnet/plamo-2-translate-eval and follow the pair-wise evaluation prompt format from its model card.

References