Outlines v1.2.0

What's Changed

• Remove dead link to VLM example from examples index by @laitifranz in #1685
• Fix: JSON schema union types (type arrays) fail with 'type must be a string' error by @brightlikethelight in #1675
• Update docs with OpenRouter guide by @Oni-giri in #1691
• Fix broken wheel builds caused by ambiguous packaging instructions by @neilmehta24 in #1690
• Support different backends on top of outlines_core by @RobinPicard in #1689
• Import fix in base.py by @Oni-giri in #1703
• Upgrade outlines-core to 0.2.11 and remove experimental CFG code by @RobinPicard in #1700

New Contributors

• @laitifranz made their first contribution in #1685
• @brightlikethelight made their first contribution in #1675
• @Oni-giri made their first contribution in #1691

Full Changelog: 1.1.1...1.2.0

Outlines v1.1.1

What's Changed

• Set add_generation_template to True in chat formatting for Transformers by @RobinPicard in #1684

Full Changelog: 1.1.0...1.1.1

Outlines v1.1.0

What's Changed

• Create the inputs module and modify the treatment of multimodal inputs by @RobinPicard in #1677
• Create the Chat input and add it to the models by @RobinPicard in #1680
• Misc doc fixes by @RobinPicard in #1683
• Remove deprecated features from v0 by @RobinPicard in #1682

Full Changelog: 1.0.4...1.1.0

Outlines v1.0.4

What's Changed

• Create Choice output type by @RobinPicard in #1671
• Add the ensure_ascii parameter to JsonSchema by @RobinPicard in #1672
• Add a batch method to models by @RobinPicard in #1674
• Fix error tokenizer arg for transformers models' generate method by @RobinPicard in #1679

Full Changelog: 1.0.3...1.0.4

Outlines 1.0.3

What's Changed

• Add llm.txt file by @rlouf in #1661
• Fix nightly tests (remove Slack alert causing troubles) by @RobinPicard in #1660
• Fix error with additionalProperties for the OpenAI model by @RobinPicard in #1668
• Add the AsyncOllama model by @RobinPicard in #1669

Full Changelog: 1.0.2...1.0.3

Outlines 1.0.2

What's Changed

• Fix typos in getting started guide by @ganglike248 in #1644
• Move VLM guide and remove Docker CI Action by @rlouf in #1643
• Fix example typo in README by @rlouf in #1650
• Add testing and documentation for vision input for VLLM model by @RobinPicard in #1648
• Add support for vision input for the Ollama model by @RobinPicard in #1647
• Add setup example for local server by @rlouf in #1646
• Add IPv4, UUID4, and Hex String Types by @jacobmarks in #1655
• Fix the FastAPI example by @rlouf in #1662

New Contributors

• @ganglike248 made their first contribution in #1644
• @jacobmarks made their first contribution in #1655

Full Changelog: 1.0.1...1.0.2

Outlines 1.0.1

What's Changed

• Fix broken links in the README by @RobinPicard in #1625
• Fix the broken links in the home page of the documentation by @RobinPicard in #1628
• Fix typos across the codebase by @sukrucildirr in #1631
• Update the documentation content and styling by @rlouf in #1636

New Contributors

• @sukrucildirr made their first contribution in #1631

Full Changelog: 1.0.0...1.0.1

Outlines v1.0.0

Why a new major version?

The v1 intends on making Outlines more closely focused on constrained generation. To do so, we delegate a wider range of tasks to the users and inference libraries. On top of making Outlines leaner, this design provides more flexibility to the users and let them use interfaces they are already familiar with.

Our approach is inspired by the unix best practices — each element does one thing well, and we compose those functional elements.

As this new version deprecates some previously available features of Outlines, we have written a migration guide that gives detailed information on how to upgrade your v0 code to v1.

Deprecated

All deprecated features listed below will be removed in version 1.1.0. Until then, a warning will be displayed with information on how to migrate your code to v1.

• The model loader functions from the models module (transformers, openai, etc.) have been deprecated. They are replaced by equivalent functions prefixed with from_ such as from_transformers, from_openai, etc. The new loader functions accept different arguments compared to the old ones. They now typically require an instance of an engine/client from the associated inference library. This change was made to avoid duplicating inference library logic and to give users more control over inference engine/client initialization.
  Documentation

# v0
from outlines import models
from transformers import BertForSequenceClassification, BertTokenizer

model = models.transformers(
    model_name="prajjwal1/bert-tiny",
    model_class=BertForSequenceClassification,
    tokenizer_class=BertTokenizer,
    model_kwargs={"use_cache": False},
    tokenizer_kwargs={"model_max_length": 512},
)

# v1
import outlines
from transformers import BertForSequenceClassification, BertTokenizer

hf_model = BertForSequenceClassification.from_pretrained("prajjwal1/bert-tiny", use_cache=False)
hf_tokenizer = BertTokenizer.from_pretrained("prajjwal1/bert-tiny", model_max_length=512)
model = outlines.from_transformers(hf_model, hf_tokenizer)

• The generate module and the associated functions (json, choice…) have been deprecated. They are replaced by the Generator constructor. While you had to select the right generate function for your output type, you can now provide any output type supported by Outlines to the unique Generator object.
  Documentation

# v0
from pydantic import BaseModel
from outlines import generate, models

class Character(BaseModel):
	name: str
		
model = models.openai("gpt-4o")
generator = generate.json(model, Character)

# v1
from openai import OpenAI
from pydantic import BaseModel
from outlines import Generator, from_openai

class Character(BaseModel):
	name: str

model = from_openai(OpenAI())
generator = Generator(model, Character)

• The TransformersVision model has been deprecated. It's replaced by TransformersMultiModal, which is more general as it supports additional input types beyond images, such as audio. When calling it, instead of providing the prompt and image assets separately, both should now be included in a single dictionary. The model is loaded with from_transformers just like the Transformers model, but the second argument must be a processor instead of a tokenizer.
  Documentation

# v0
from io import BytesIO
from urllib.request import urlopen
from PIL import Image
from transformers import LlavaForConditionalGeneration
from outlines import models, generate

def img_from_url(url):
    img_byte_stream = BytesIO(urlopen(url).read())
    return Image.open(img_byte_stream).convert("RGB")

model = models.transformers_vision(
    model_name="trl-internal-testing/tiny-LlavaForConditionalGeneration",
    model_class=LlavaForConditionalGeneration,
)
generator = generate.text(model)
result = generator(
    "Describe the image <image>",
    img_from_url("https://upload.wikimedia.org/wikipedia/commons/2/25/Siam_lilacpoint.jpg")
)

# v1
from io import BytesIO
from urllib.request import urlopen
from PIL import Image
from transformers import LlavaForConditionalGeneration, AutoProcessor
import outlines

def img_from_url(url):
    img_byte_stream = BytesIO(urlopen(url).read())
    return Image.open(img_byte_stream).convert("RGB")

model = outlines.from_transformers(
	LlavaForConditionalGeneration.from_pretrained("trl-internal-testing/tiny-LlavaForConditionalGeneration"),
	AutoProcessor.from_pretrained("trl-internal-testing/tiny-LlavaForConditionalGeneration")
)
image = img_from_url("https://upload.wikimedia.org/wikipedia/commons/2/25/Siam_lilacpoint.jpg")
result = model({"text": "Describe the image <image>", "images": image})

• The Exllamav2 model has been deprecated without replacement because its interface is not fully compatible with Outlines. We had to implement cumbersome patching to make it work, so we decided to remove it entirely.

• The function module and the associated Function class have been deprecated. They are replaced by the Application class, which serves a similar purpose to Function. There are two notable differences: an Application is not initialized with a model (a model must be provided when calling the object), and template variables must be provided in a dictionary instead of as keyword arguments when calling the Application.
  Documentation

# v0
from pydantic import BaseModel
from outlines import Function, Template

class Character(BaseModel):
	name: str

template = Template.from_string("Create a {{ gender }} character.")
fn = Function(template, Character, "hf-internal-testing/tiny-random-GPTJForCausalLM")
response = fn(gender="female")

# v1
from pydantic import BaseModel
from outlines import Application, Template, from_transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

class Character(BaseModel):
	name: str

model = from_transformers(
    AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct"),
    AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
)

template = Template.from_string("Create a {{ gender }} character.")
app = Application(template, Character)
response = app(model, {"gender": "female"})

• The samplers module and the associated objects (multinomial, greedy…) have been deprecated. You should now use the inference arguments specific to the inference library your model is based on to control the sampling.

# v0
from outlines import generate, models, samplers

model = models.transformers("microsoft/Phi-3-mini-4k-instruct")
generator = generate.text(model, samplers.beam_search(2))
response = generator("Write a short story about a cat", max_tokens=10)

# v1
from outlines import Generator, from_transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model = from_transformers(
    AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct"),
    AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
)
response = model("Write a short story about a cat", num_beams=2)

• The load_lora methods on the VLLM and LlamaCpp models have been deprecated. You should now load through the Llama instance provided when initializing the model in the case of the LlamaCpp model, and provide it as a keyword argument when calling the model in the case of the VLLM model.

# v0
from outlines import models
from vllm import LLM

model = models.vllm("erwanf/gpt2-mini")
model.load_lora("path/to/lora/file")
response = model("Write a short story about a cat.")

#v1
from outlines import from_vllm
from vllm import LLM
from vllm.lora.request import LoRARequest

model = from_vllm(
    LLM("microsoft/Phi-3-mini-4k-instruct")
)
lora_request = LoRARequest("path/to/lora/file", 1, "path/to/lora/file")
response = model("Write a short story about a cat.", lora_request=lora_request)

Modified

Some objects are maintained, but their interface or behavior has been modified.

• The interface of Model classes (Transformers, OpenAI, etc.) has been significantly modified. Models can now be called directly with a prompt and an output type without having to create a generator first. Additionally, all models have a stream method that can be invoked directly by the user.
  Documentation

# v0
from pydantic import BaseModel
from outlines import generate, models

class Character(BaseModel):
		name: str

model = models.openai("gpt-4o")
generator = generate.json(model, Character)
result = generator("Create a character")

# v1
from openai import OpenAI
from pydantic import BaseModel
from outlines import from_openai

class Character(BaseModel):
	name: str

model = from_openai(OpenAI(), "gpt-4o")
result = model("Create a character", Character)

• The interface of the __init__ method of the OpenAI model class has been modified. While it previously accepted a client and an OpenAIConfig object instance, it now accepts a client and a model name. The inference arguments from the config object should now be specified when calling the model to more closely align with the OpenAI Python library's functionality. If you provide an OpenAIConfig instance when initializing the model, a deprecation warning will appear and your model will behave like a v0 model.
  We recommend using the from_openai function instead of initializing models directly.
  [Documentation](https://dottxt-ai.gith...

Outlines v0.2.3

What's Changed

• JSON schema error when using OpenAI with Pydantic model by @derfred in #1472
• Tweak version commands in bug_report.yml by @nchammas in #1476
• Fix capitalization in installation.md by @cpfiffer in #1488
• Remove all issue templates except bug reports by @rlouf in #1496
• Update the logo by @rlouf in #1521

New Contributors

• @nchammas made their first contribution in #1476

Full Changelog: 0.2.2...0.2.3

Outlines v0.2.2

What's Changed

• Update prompt templating documentation by @rlouf in #1448
• Format files and fix numerous mypy warnings, thanks to ruff by @yvan-sraka in #1470
• Remove pre-commit runtime dependency by @yvan-sraka in #1468
• Improve the regex DSL (replace | by either and introduce new quantifiers) by @rlouf in #1463

Full Changelog: 0.2.1...0.2.2