metadata

license: apache-2.0
language:
  - en
tags:
  - ctranslate2
  - fastchat-t5-3b
  - quantization
  - int8
pipeline_tag: text2text-generation

Model Card for FastChat-T5 3B Q8

The model is quantized version of the lmsys/fastchat-t5-3b-v1.0 with int8 quantization.

Model Details

Model Description

The model being quantized using CTranslate2 with the following command:

ct2-transformers-converter --model lmsys/fastchat-t5-3b --output_dir lmsys/fastchat-t5-3b-ct2 --copy_files generation_config.json added_tokens.json tokenizer_config.json special_tokens_map.json spiece.model --quantization int8 --force --low_cpu_mem_usage

If you want to perform the quantization yourself, you need to install the following dependencies:

pip install -qU ctranslate2 transformers[torch] sentencepiece accelerate

Shared by: Lim Chee Kin
License: Apache 2.0

How to Get Started with the Model

Use the code below to get started with the model.

import ctranslate2
import transformers

translator = ctranslate2.Translator("limcheekin/fastchat-t5-3b-ct2")
tokenizer = transformers.AutoTokenizer.from_pretrained("limcheekin/fastchat-t5-3b-ct2")

input_text = "translate English to German: The house is wonderful."
input_tokens = tokenizer.convert_ids_to_tokens(tokenizer.encode(input_text))

results = translator.translate_batch([input_tokens])

output_tokens = results[0].hypotheses[0]
output_text = tokenizer.decode(tokenizer.convert_tokens_to_ids(output_tokens))

print(output_text)

The code is taken from https://opennmt.net/CTranslate2/guides/transformers.html#t5.

The key method of the code above is translate_batch, you can find out its supported parameters here.