Skip to content

transformers

Installation

You need to install the transformer, datasets and torch libraries to be able to use these models in Outlines:

pip install torch transformers datasets

Outlines provides an integration with the torch implementation of causal models in the transformers library. You can initialize the model by passing its name:

from outlines import models

model = models.transformers("microsoft/Phi-3-mini-4k-instruct", device="cuda")

If you need more fine-grained control you can also initialize the model and tokenizer separately:

from transformers import AutoModelForCausalLM, AutoTokenizer
from outlines import models

llm = AutoModelForCausalLM.from_pretrained("gpt2", output_attentions=True)
tokenizer = AutoTokenizer.from_pretrained("gpt2")
model = models.Transformers(llm, tokenizer)

Using Logits Processors

There are two ways to use Outlines Structured Generation with HuggingFace Transformers:

  1. Use Outlines generation wrapper, outlines.models.transformers
  2. Use OutlinesLogitsProcessor with transformers.AutoModelForCausalLM

Outlines supports a myriad of logits processors for structured generation. In these example, we will use the RegexLogitsProcessor which guarantees generated text matches the specified pattern.

Using outlines.models.transformers

import outlines

time_regex_pattern = r"(0?[1-9]|1[0-2]):[0-5]\d\s?(am|pm)?"

model = outlines.models.transformers("microsoft/Phi-3-mini-4k-instruct", device="cuda")
generator = outlines.generate.regex(model, time_regex_pattern)

output = generator("The the best time to visit a dentist is at ")
print(output)
# 2:30 pm

Using models initialized via the transformers library

import outlines
import transformers


model_uri = "microsoft/Phi-3-mini-4k-instruct"

outlines_tokenizer = outlines.models.TransformerTokenizer(
    transformers.AutoTokenizer.from_pretrained(model_uri)
)
phone_number_logits_processor = outlines.processors.RegexLogitsProcessor(
    "\\+?[1-9][0-9]{7,14}",  # phone number pattern
    outlines_tokenizer,
)

generator = transformers.pipeline('text-generation', model=model_uri)

output = generator(
    "Jenny gave me her number it's ",
    logits_processor=transformers.LogitsProcessorList([phone_number_logits_processor])
)
print(output)
# [{'generated_text': "Jenny gave me her number it's 2125550182"}]
# not quite 8675309 what we expected, but it is a valid phone number

Alternative Model Classes

outlines.models.transformers defaults to transformers.AutoModelForCausalLM, which is the appropriate class for most standard large language models, including Llama 3, Mistral, Phi-3, etc.

However other variants with unique behavior can be used as well by passing the appropriate class.

Mamba

Mamba is a transformers alternative which employs memory efficient, linear-time decoding.

To use Mamba with outlines you must first install the necessary requirements:

pip install causal-conv1d>=1.2.0 mamba-ssm torch transformers

Then you can either create an Mamba-2 Outlines model via

import outlines

model = outlines.models.mamba("state-spaces/mamba-2.8b-hf")

or explicitly with

import outlines
from transformers import MambaForCausalLM

model = outlines.models.transformers(
    "state-spaces/mamba-2.8b-hf",
    model_class=MambaForCausalLM
)

Read transformers's documentation for more information.

Encoder-Decoder Models

You can use encoder-decoder (seq2seq) models like T5 and BART with Outlines.

Be cautious with model selection though, some models such as t5-base don't include certain characters ({) and you may get an error when trying to perform structured generation.

T5 Example:

import outlines
from transformers import AutoModelForSeq2SeqLM

model_pile_t5 = models.transformers(
    model_name="EleutherAI/pile-t5-large",
    model_class=AutoModelForSeq2SeqLM,
)

Bart Example:

model_bart = models.transformers(
    model_name="facebook/bart-large",
    model_class=AutoModelForSeq2SeqLM,
)