Changelog

This package is already used in production at Tune AI, please do not wait for release 1.x.x for stability, or expect to reach 1.0.0. We do not follow the general rules of semantic versioning, and there can be breaking changes between minor versions.

All relevant steps to be taken will be mentioned here.

8.0.3

  • Fix bug in tools that was causing ever increasing number of tools in the Thread object

  • OpenAI protocol abstracted away as a OpenAIProtocol class in tuneapi.apis.openai. This is to make it easier to add new endpoints in the future.

8.0.2

  • Added usage tracking for OpenAI and Anthropic

8.0.1

  • Typo so now we are in 8.x.x series

  • Fix bug in structured generation for Openai.

0.8.0

  • Adding multimodal support for OpenAI and Gemini endpoints.

  • Added new types for the new endpoints, etc.

  • New utils function file_size to get the size of the file in bytes.

0.7.4

  • Remove model_tune.py, model_groq.py and model_mistral.py and subclass those models from Openai.

0.7.3

  • Fix bug in image modality for Gemini

  • Fix bug in structured generation for OpenAI by setting strict=True in the API.

  • Fix bug in structured generation for Gemini

  • Abstracted making input structure for Anthropic, OpenAI and Gemini using _process_input method

  • Abstracted outputs for Anthropic, OpenAI and Gemini using _process_output method

0.7.2

  • Add image modality for Gemini and OpenAI to complement Anthropic.

0.7.1

  • Add structured genration support for Gemini and OpenAI APIs. You can jsut pass schema to Thread. model.chat will take care of it automatically. Here’s an example:

    from tuneapi import tt, ta
from pydantic import BaseModel
from typing import List, Optional, Dict, Any

class MedicalRecord(BaseModel):
    date: str
    diagnosis: str
    treatment: str

class Dog(BaseModel):
    name: str
    breed: str
    records: Optional[List[MedicalRecord]] = None

class Dogs(BaseModel):
    dogs: List[Dog]

model = ta.Gemini()
out: Dogs = model.chat(tt.Thread(
    tt.human("""
        At the Sunny Paws Animal Clinic, we keep detailed records of all our furry patients. Today, we saw a few dogs.
        There was 'Buddy,' a golden retriever, who visited on '2023-10-26' and was diagnosed with a 'mild ear infection,'
        which we treated with 'ear drops.' Then, there was 'Luna,' a playful beagle, who came in on '2023-10-25' for a
        'routine check-up,' and no treatment was needed, but we also had her back on '2023-10-28' with a 'upset tummy'
        which we treated with 'bland diet and probiotics.' Finally, a third dog named 'Rocky', a small terrier mix,
        showed up on '2023-10-29' with a small 'cut on his paw,' we cleaned it and used an 'antibiotic ointment'. We
        also have 'Daisy,' a fluffy poodle, who doesn't have any medical records yet, thankfully!
    """),
    schema=Dogs,
))

for dog in out.dogs:
    print(f"Dog: {dog.name}, Breed: {dog.breed}")
    if dog.records:
        for record in dog.records:
            print(f"  Date: {record.date}, Diagnosis: {record.diagnosis}, Treatment: {record.treatment}")
    else:
        print("  No medical records on file.")
    print()

  • Add pydantic as a dependency in the package.

0.7.0

  • All models now have <model>.distributed_chat_async that can be used in servers without blocking the main event loop. This will give a much needed UX improvement to the entire system.

0.6.3

  • <model>.distributed_chat now takes in args that are passed to the post_logic.

0.6.2

  • New set of utils in tuneapi.utils called prompt to help with the basics of prompting.

0.6.1

  • Package now uses fire==0.7.0

0.6.0

  • distributed_chat functionality in tuneapi.apis.turbo support. In all APIs search for model.distributed_chat() method. This enables fault tolerant LLM API calls.

  • Moved tuneapi.types.experimental to tuneapi.types.evals

0.5.13

  • tuneapi.types.ModelInterface has an extra_headers attribute in it.

0.5.12

  • Remove code to sanitize assistant message in for Tune and OpenAI LLM APIs.

0.5.11

  • Fix bug where parallel_tool_calls was sent even for non tool calls.

0.5.10

  • Remove redundant prints.

0.5.9

  • By default set the value parallel_tool_calls in OpenAI to False.

0.5.8

  • If you have numpy installed in your environment, then tuneapi.utils.randomness.reservoir_sampling will honour the seed value. If you do not have numpy installed, then the seed value will be ignored.

  • Fix Bug in Gemini API body for functions with no parameters.

0.5.7

  • Implement extra_headers via __init__ as well.

0.5.6

  • Remove protobuf as a dependency in because bunch of other packages break. The functions are still present

0.5.5

  • In all implmenetations of tuneapi.types.chats.ModelInterface add new input to the API endpoints called extra_headers which is a dictionary to update the outgoing headers.

0.5.4

  • Standardise tuneapi.types.chats.ModelInterface to have model_id, api_token added to the base class.

0.5.3

  • Fix bug in Tune proxy API where incorrect variable stop_sequence was sent instead of the correct stop causing incorrect behaviour.

  • bump dependency to protobuf>=5.27.3

  • remove __version__ from tuneapi package

  • remove CLI entrypoint in pyproject.toml

0.5.2

  • Add ability to upload any file using tuneapi.endpoints.FinetuningAPI.upload_dataset_file to support the existing way to uploading using threads.

0.5.1

  • Fix bug in the endpoints module where error was raised despite correct inputs

0.5.0 (breaking)

In this release we have moved all the Tune Studio specific API out of tuneapi.apis to tuneapi.endpoints to avoid cluttering the apis namespace.

- from tuneapi import apis as ta
+ from tuneapi import endpoints as te
...
- ta.ThreadsAPI(...)
+ te.ThreadsAPI(...)

  • Add support for finetuning APIs with tuneapi.endpoints.FinetuningAPI

  • Primary environment variables have been changed from TUNE_API_KEY to TUNEAPI_TOKEN and from TUNE_ORG_ID to TUNEORG_ID, if you were using these please update your environment variables

  • Removed CLI methods test_models and benchmark_models, if you want to use those, please copy the code from this commit

0.4.18

  • Fix bug where function response was tried to be deserialised to the JSON and then sent to the different APIs.

0.4.17

  • Fix error in tuneapi.utils.serdeser.to_s3 function where content type key was incorrect

0.4.16

  • Adding support for python 3.12

  • Adding tool as a valid role in tuneapi.types.chats.Message

0.4.15

  • When there is an error in the model API, we used to print the error message. Now we are returning the error message in the response.

0.4.14

  • Fix bug where a loose pydantic import was present

0.4.13

  • Bug fixes in JSON deserialisation

0.4.12

  • Fix bug in Threads API where incorrect structure was sent by client

  • Add images support for Anthropic API

  • Add Message.images field to store all images